# An AI Research Agent That Runs Your Experiments — Tim Sweeney, Weights & Biases Tim Sweeney, Weights & Biases AI Engineer World's Fair 2026 [Talk page](https://ai.engineer/talks/hd7TOvmyAxU-ai-research-agent-that-runs-your-experiments) · [Video](https://www.youtube.com/watch?v=hd7TOvmyAxU) ## An AI Research Agent That Runs Your Experiments ARIA turns a research request into GPU jobs, experiment analysis, and visual reports inside Weights & Biases. A live batch shows how that loop works; the team's own traces and nightly evaluations show how they decide whether the agent itself is improving. ### Start with a project that already has 200 experiments [0:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=13s) A research workspace can accumulate experiments faster than a person can explain them. Tim Sweeney, Principal Engineer at Weights & Biases by CoreWeave, opens a project with more than 200 training jobs. A scatter plot shows loss declining over time. The next problem is deciding what those runs teach—and what experiment to run next. ![Source frame: Start with a project that already has 200 experiments](https://ai.engineer/api/talk-essay-media/cb6f760dd8963577108b44f7cbd1657533a58a052645d8ba1c5bdcb8542bb4ad) Source frame: Start with a project that already has 200 experiments The project uses Karpathy's autoresearch, a small codebase that trains an LLM and provides a manageable setting for repeated changes. **ARIA**, the AI research and iteration agent, opens as a chat inside the workspace. Project context and images can accompany a request, so the conversation can begin with the experiment data already close at hand. The useful starting point is a long-running conversation rather than the agent's onstage introduction. In that conversation, ARIA has helped download the code, configure Launch jobs, and set up GPUs. It can then iterate on both the code and its hyperparameters. Those are different kinds of changes: one alters the implementation being trained; the other changes how that implementation trains. ### A chat request becomes a batch of GPU jobs [4:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=270s) Sweeney asks ARIA to conduct another batch of experiments and try to find the best model live. He adds encouragement—because, naturally, the models need encouragement. The consequential decision comes next: ARIA avoids a large architecture change, treating it as too risky for this iteration. Sweeney expects it to try hyperparameter modifications instead, and a shell call starts the experimentation loop. ![Source frame: A chat request becomes a batch of GPU jobs](https://ai.engineer/api/talk-essay-media/5506d386e00079d9d0be15d1a29e3e10b8bf015d27e4a94161100d3fda314af4) Source frame: A chat request becomes a batch of GPU jobs W&B Launch supplies the connection to compute. A Launch queue lets humans or agents submit long-running experiment jobs to a cluster, including jobs that need GPUs. Sweeney opens terminal output from his Kubernetes cluster to show experiments executing, then returns to ARIA, which is polling while it waits for the work to finish. The chat initiates and monitors the work; the cluster performs the training. Where does the time-consuming work happen after a research request? The flow below separates the agent's decision and shell execution from queued GPU training. The polling path makes the separation visible: launching a job does not mean its result is immediately available. From request to running experiments ARIA starts the experiment loop, Launch connects it to GPU compute, and the agent polls while training proceeds. - Research request — Conduct another batch and look for a better model. - ARIA — Chooses an experiment direction and starts a shell call. - W&B Launch queue — Connects submitted experiment jobs to compute. - Kubernetes GPU cluster — Runs the training experiments. - Research request → ARIA (Starts a new iteration) - ARIA → W&B Launch queue (Starts experiment jobs) - W&B Launch queue → Kubernetes GPU cluster (Launches on compute) - ARIA → W&B Launch queue (Polls for completion) ### Turn accumulated runs into explanations and charts [5:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=355s) While training runs, a second request asks for a summary of the project's highest-performing runs. This is useful when a teammate joins a project or returns from time away: the experiment history needs to become an explanation of what worked. A previously prepared pattern-finding example goes further, identifying an emerging model family, batch size as an influential parameter, and a promising architectural recipe. These are leads for further research, rather than a demonstrated guarantee that any one change causes an improvement. ![Source frame: Turn accumulated runs into explanations and charts](https://ai.engineer/api/talk-essay-media/f84e207bf5732f68c0bf6fd778def722f523546b017476e577e3464ccfb9d751) Source frame: Turn accumulated runs into explanations and charts The analysis can become a native W&B artifact rather than remaining a chat answer: - **Reports:** A report combines Markdown with embedded plots, charts, and graphics—Sweeney calls it a “markdown file on steroids.” The example includes a project thesis, data panels, and a parameter-importance chart showing relationships among parameters. - **Workspaces:** ARIA is tuned and prompted to construct workspaces and plots using W&B's built-in chart types. That gives the analysis a form colleagues can inspect in the interface they already use. A check-in shows the summary request querying W&B, applying patches, and writing code, while the training conversation continues polling for results. Analysis and training therefore have their own work to complete. ARIA's value here comes from connecting research data, executable tools, and visualization primitives within the same product. The iOS announcement extends that workflow to a phone: researchers can steer hyperparameter-tuning jobs while away from their desks. The larger ambition is an end-to-end automated research platform that handles job orchestration, GPU workloads, research lookup, and hypothesis collaboration. Sweeney frames the division of labor around letting ARIA handle mechanics while researchers concentrate on new ideas, architectures, and parameters. ### Separate conversation handling from long-running compute [10:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=612s) ARIA's backend begins with a familiar sequence: web and iOS clients communicate with an API server, the server writes to a turn database, and a worker harness processes the work. The harness then connects to the capabilities needed to carry out a research request. ![Source frame: Separate conversation handling from long-running compute](https://ai.engineer/api/talk-essay-media/d05ca504df96f2e4c689a30aed23dd6a8b913f66730f0927dc9a5e0996d08113) Source frame: Separate conversation handling from long-running compute Those connections serve distinct purposes: - **Sandbox execution:** Shell calls and Python data-science work give the agent a way to manipulate code and analyze data. - **Model inference:** An LLM provider supplies the model calls used by the worker; W&B Inference is presented as one option. - **External training:** Launch handles workloads outside the agent's main loop, including training that can take days, with CoreWeave GPUs supplying compute. - **Observability:** Sessions, turns, tool calls, and errors are logged so the team can inspect what the agent actually did. Sweeney says ARIA logs 100% of its traces to W&B Weave. The trace store also feeds development. Offline work turns observed behavior into tasks, evaluates candidate agents, and records results in a shared dashboard. One loop improves the tests; another forms hypotheses, implements candidates, and studies their evaluation results. The loops are complementary but also adversarial: better tests can expose weaknesses in a candidate that previously looked promising. The selected agent returns to production through a registry. ### Read the shape of a conversation, then inspect its failures [12:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=779s) The [Weave](https://wandb.ai/site/weave) dashboard gives a bird's-eye view through span volume, conversation volume, and token tracking. Sweeney then opens a conversation feed filtered to internal employees. Its span view makes trace topology visible: different colors and shapes distinguish tool calls, LLM calls, and thinking blocks. Opening a conversation reveals the system prompt, user messages, shell calls, and reasoning blocks behind that shape. ![Source frame: Read the shape of a conversation, then inspect its failures](https://ai.engineer/api/talk-essay-media/e5d7ca0df5253605519d7df3986db52a43587e762a3daa45de2da0e839b89079) Source frame: Read the shape of a conversation, then inspect its failures That detailed view is where the research lead, engineer, and product manager add notes and feedback before turning behavioral discoveries into tasks. A contextual Summarize button can also start an ARIA conversation about the item being inspected. In this example, ARIA analyzes one of its own conversations to recommend improvements to ARIA. The agent assists the review, while the team still has the underlying trace available. Signals add automated triage to live traffic. LLM judges flag user frustration, low-quality responses, and other behaviors, helping the team find clusters to address in a later iteration. One displayed judge identifies frustration because the user expresses dissatisfaction with a loss curve. Its reasoning is inspectable, which gives reviewers a concrete explanation to assess rather than just a flag. The flag identifies dissatisfaction; it does not by itself establish what caused the poor experience. ### Turn observed behavior into nightly tests [16:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=962s) A task is a YAML file that acts as a unit test for the agent. The demonstrated prompt asks ARIA to compare two runs that both give good results: what differs, and what can be learned? The task pairs that request with definitions of success. This makes a broad product expectation—give useful research advice—specific enough to evaluate repeatedly. ![Source frame: Turn observed behavior into nightly tests](https://ai.engineer/api/talk-essay-media/987eb4361c2cad7a72f9b4496fde8e60575d8c54c93d09778e9c8683388e8032) Source frame: Turn observed behavior into nightly tests The example checks three separate properties: - **Correctness:** An LLM judge evaluates the answer against correctness defined for this particular question. - **Interesting insights:** A second LLM judge asks whether the comparison produces insights worth knowing. A correct answer can still be unhelpful. - **Expediency:** A rule-based judge checks whether the agent produces a result within six tool calls, giving efficiency an explicit test alongside answer quality. About 200 tasks form a suite that runs nightly, with results tracked in Weave. The displayed candidate scores 73%, compared with the production model's 72%; Sweeney says the team plans to move it forward that Friday. These scores describe performance on the team's suite. The recording does not establish whether the one-percentage-point difference is statistically significant or predicts improvement outside that suite. How does a production conversation influence the next release? The cycle below shows the intermediate steps: review produces tasks, tasks test candidates, and the results inform a team decision. Production traces provide the starting material; they become useful release evidence only after the team defines the behavior it wants to test. The agent improvement loop Live behavior supplies examples; explicit tasks and candidate evaluations turn those examples into release decisions. - Production traces — Collect conversations and agent actions. - Behavior review — Humans, ARIA, and LLM judges identify useful patterns and failures. - YAML tasks — Define prompts and checks for desired behavior. - Candidate evaluations — Run the nightly suite and compare results in Weave. - Team release decision — Use shared results to select and promote an agent. - Production traces → Behavior review (Inspect behavior) - Behavior review → YAML tasks (Turn findings into tests) - YAML tasks → Candidate evaluations (Test candidates nightly) - Candidate evaluations → Team release decision (Compare and decide) - Team release decision → Production traces (Return the selected agent to production) ### Keep humans in the loop—and let the experiment finish [18:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1123s) The production advice follows directly from this development loop: - **Instrument behavior:** Log sessions, turns, tools, and feedback to catch behavioral bugs. An agent can complete its calls without producing the behavior a user needs. - **Use evaluations for decisions:** Researchers and engineers should develop tasks together and treat performance metrics as go/no-go evidence. - **Keep human judgment:** Use the product and review its best and worst traces as a team. LLM judges miss behavioral nuances, so automated scores cannot cover the whole review. - **Improve context and tools:** Before adding elaborate harness or memory machinery, give the agent the business domain, available primitives, and relevant business data. Sweeney's team found substantial opportunities in those simpler additions. ![Source frame: Keep humans in the loop—and let the experiment finish](https://ai.engineer/api/talk-essay-media/1fd76f1de783c381dd91006cbd496a3237ff828107156b8ff121411f3d8d4d0d) Source frame: Keep humans in the loop—and let the experiment finish The final check returns to the batch started onstage. A new point has appeared in the workspace, and the batch has run 12 experiments. The reported result is 5.833, close to the earlier best but without beating it. ARIA has carried a request through experiment selection, shell execution, queued GPU training, waiting, and a visible result. The automation completed a research iteration; an improvement remained something the experiments had to earn. ## Key takeaways - Keep long-running training outside the agent's main loop: ARIA starts experiments through Launch and polls while GPU jobs execute. - Evaluate useful research behavior along separate dimensions: the example task checks correctness, interesting insights, and a six-tool-call limit. - Production traces become more useful when human review and live judges turn observed behavior into repeatable tasks and candidate evaluations. - Domain context and available tools are practical places to improve an agent before adding elaborate harness or memory machinery. - The live batch completed 12 experiments and nearly matched the earlier best result; autonomous execution did not guarantee a better model. ## Resources - [W&B Weave](https://wandb.ai/site/weave) — The tracing and evaluation product used in the talk to inspect conversations, run live judge signals, and compare nightly candidate results. ## Slides ### ARIA introduction Sweeney asks ARIA to introduce itself. ARIA describes analyzing experiments and says it is looking at a filtered workspace of runs logging val/loss, grouped by iteration_group. [3:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=236s) [The assistant panel shows a submitted introduction prompt and a loading indicator.](https://assets.ai.engineer/sites/aiecode2025/sha256/78eb74a0b1ebb4a39dbe6bda90ac46d8a8d05a5f4d7dc9565d660fc797bab322.webp) ARIA Hello ARIA, you are on stage at AIEngineer Worlds Fair 2026! Please introduce yourself! Ask ARIA Interface text: Tsweeney's workspace Personal workspace Runs 248 Search runs 1 filter NAME NOTES ITEF VAL_B Hypothesis: Width7... worlds-f[illegible] 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity[illegible] 6.03687 Hypothesis: Width5... width51[illegible] 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor 6.18444 Hypothesis: The lo[illegible] 1-23 of 23 Y-axis ticks: 11.5, 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5, 7, 6.5, 6 Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney [4:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=244s) [ARIA introduces itself and describes analyzing experiments in a W&B workspace.](https://assets.ai.engineer/sites/aiecode2025/sha256/ec3a2abcafe92b5c13f0a65e7e0462a649129367939f850d24ce98d3be1ddf0b.webp) ARIA Hello ARIA, you are on stage at AIEngineer Worlds Fair 2026! Please introduce yourself! Hello AI Engineer Worlds Fair 2026! 👋 I’m ARIA — the AI Research and Iteration Agent inside Weights & Biases. I help researchers and engineers move faster from “what happened?” to “what should we try next?” I can inspect W&B projects, compare experiment runs, analyze metrics and configs, debug failed or slow jobs, summarize Weave traces and evaluations, create reports or workspace panels, and even help launch new bounded experiment iterations when a project is set up for it. Right now I’m looking at a W&B project with hundreds of runs, including a filtered workspace focused on runs that logged val/loss, grouped by iteration_group. That means I can do things like: • identify the best-performing run or group, • compare validation loss trends across iterations Ask ARIA Interface text: Tsweeney's workspace Personal workspace Runs 248 Search runs 1 filter NAME NOTES ITEF VAL_B Hypothesis: Width7... worlds-f[illegible] 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity[illegible] 6.03687 Hypothesis: Width5... width51[illegible] 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor 6.18444 Hypothesis: The lo[illegible] 1-23 of 23 Y-axis ticks: 11.5, 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5, 7, 6.5, 6 Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney ### Best model and next steps Sweeney opens a long-running ARIA chat from over 200 auto-research experiments. ARIA reports the best run at 5.749383578384566 val_bpb and recommends exploring near width640 with 2400 steps. [4:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=253s) [ARIA identifies the current best model and recommends settings for further experiments.](https://assets.ai.engineer/sites/aiecode2025/sha256/2e9087f42a3c2959edcf93465fa0afd00b802b823e84ac452e170f76ba868960.webp) • tiny LR changes hurt: • 0.00095 → 5.8428 • 0.00105 → 5.8544 • pushing to seq896 was not enough to win: 5.7946 Best model remains run: rk6vjvz4 variant: WFF02_s768_w640_b28_2400 val_bpb: 5.749383578384566 The live encore didn’t break the record, but it gave us a much sharper map: stay near width640, lr=0.001, floor=0.03, 2400 steps, and explore stochastic repeats / seq800–832 carefully. Interface text: Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:34:13 wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb Tsweeney's workspace Personal workspace Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts Runs 248 Search runs 1 filter NAME NOTES ITE[illegible] VAL_B Hypothesis: Width7... worlds-f[illegible] 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity[illegible] 6.03687 Hypothesis: Width5... width51[illegible] 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor 6.18444 Hypothesis: The lo[illegible] 1-23 of 23 Chart Y-axis ticks: 11.5, 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5, 7, 6.5, 6 Chart X-axis tick: 0 ARIA Ask ARIA ### Experiment comparisons ARIA compares nearby runs and finds that changes to batch size, learning rate and steps did not beat the leading model. Sweeney says the long-running chat has used the auto research loop for over 200 experiments. [4:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=261s) [ARIA lists nearby experiment results and shows the current best model.](https://assets.ai.engineer/sites/aiecode2025/sha256/4184f2f6035250c21a6187423e884e00657a7826015f2f82fb6d0e03f00fec69.webp) The encore ruled out several nearby ideas: • batch26 at seq768 was bad: 5.8901 • batch30 at seq768 was worse than batch28: 5.8577 • 2350 and 2450 steps were worse than 2400 • tiny LR changes hurt: • 0.00095 → 5.8428 • 0.00105 → 5.8544 • pushing to seq896 was not enough to win: 5.7946 Best model remains run: rk6vjvz4 variant: WFF02_s768_w640_b28_2400 val_bpb: 5.749383578384566 Interface text: Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:34:22 wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb Tsweeney's workspace Personal workspace Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts Runs 248 Search runs 1 filter NAME NOTES ITEF VAL_B Hypothesis: Width7... worlds-f... 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity... 6.03687 Hypothesis: Width5... width51... 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor... 6.18444 Hypothesis: The lo... 1-23 of 23 Y-axis ticks: 11.5, 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5, 7, 6.5, 6 ARIA Ask ARIA ### Choosing a new experiment batch ARIA selects three leading configurations for a 12-run seedstorm, reasoning that variance within the winning configuration range could produce a new best result. [5:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=309s) [ARIA lists three leading configurations and proposes a 12-run batch.](https://assets.ai.engineer/sites/aiecode2025/sha256/49512cdfaaa30cdbc7a2df395719d4618f53d238cf4e15349575abbd3f8da333.webp) ARIA I’ll launch another 12-run Worlds Fair batch, weighted toward champion repeats and the closest runner-up shape. shell python - <<'PY' import wandb, json api=wandb.Api(timeout=120) path='aria-autoresearch-01/wandb_agent... Thinking Creating a script for seedstorm Confirmed. Ignoring the toy smoke runs, the frontier top is still: • rk6vjvz4 — seq768/w640/b28/2400 — 5.7494 • xjgxqrx[illegible] — seq832/w640/b28/2400 — 5.7551 • xnlnckfm — seq800/w640/b28/2400 — 5.7610 I’m launching a seedstorm wave: 12 runs concentrated on these three shapes, because the data says variance inside the winning basin is large enough that a new best can plausibly come from another favorable trajectory. Ask ARIA Interface text: Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:35:09 wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products aria-autoresearch-01 Projects wandb_agent_default_pr... Workspace Tim Sweeney wandb Tsweeney's workspace Personal workspace Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts Runs 248 Search runs 1 filter NAME NOTES ITEF VAL_B Hypothesis: Width7... worlds-f... 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity... 6.03687 Hypothesis: Width5... width512... 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor 6.18444 Hypothesis: The lo... 1–23 of 23 Y-axis ticks: 11.5, 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5, 7, 6.5, 6 X-axis tick: 0 ### Launch queue Sweeney shows a Kubernetes Launch queue with finished runs. He explains that Launch connects compute clusters so people and agents can run experimentation jobs. [5:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=318s) [A Kubernetes Launch queue shows a table of finished runs.](https://assets.ai.engineer/sites/aiecode2025/sha256/b9a49035bf5b373e9846e79b5e868c4cf4b6f07758e2d9abb59738b900a9b8ab.webp) Launch autoresearch-k8s aria-autoresearch-01 Queue Kubernetes Active Prioritization Runs Agents Config Monitor Completed and failed runs automatically clear after 30 days. Search 1–20 of 100+ Run | Job | Project | Priority | Status | Errors | Author | Agent | Submitted ... | Start time | Duration wfe-dock... | | wandb_ager | Medium | Finished | | tsweeney | cdek997f | Jul 1st 2026 at 11:58am | Jul 1st 2026 at 12:02pm | 1m 30s wfe-dock... | | wandb_ager | Medium | Finished | | tsweeney | cdek997f | Jul 1st 2026 at 11:58am | Jul 1st 2026 at 12:02pm | 1m 37s wfe-dock... | | wandb_ager | Medium | Finished | | tsweeney | cdek997f | Jul 1st 2026 at 11:58am | Jul 1st 2026 at 12:02pm | 1m 38s wfe-dock... | | wandb_ager | Medium | Finished | | tsweeney | cdek997f | Jul 1st 2026 at 11:58am | Jul 1st 2026 at 12:02pm | 1m 29s wfe-dock... | | wandb_ager | Medium | Finished | | tsweeney | cdek997f | Jul 1st 2026 at 11:58am | Jul 1st 2026 at 12:02pm | 1m 30s Interface text: aria-autoresearch-01 Launch Tim Sweeney wandb Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:35:22 wandb.ai/aria-autoresearch-01/launch/[illegible] D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products Ask Gemini Work ### Experiment batch submitted ARIA reports that its 12-run seedstorm is live, adds a run filter, and polls for results against the current record of 5.7493835784. [5:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=345s) [ARIA reports that the 12-run batch is live and displays a run filter and current record.](https://assets.ai.engineer/sites/aiecode2025/sha256/bff040683cccc3bad162391bb2d11f0deb7b33b2488acf08f3c9568ae94a960d.webp) ARIA I’m launching a seedstorm wave: 12 runs concentrated on these three shapes, because the data says variance inside the winning basin is large enough that a new best can plausibly come from another favorable trajectory. apply_patch /workspace/submit_worlds_fair_seedstorm_12_trainpy_docker_k8s.py Seedstorm script is ready. Submitting all 12 to the K8s Launch queue now. shell python -m py_compile /workspace/train_rope_swiglu_b128_frontier.py /workspace/submit_worlds_fair_se... The 12-run seedstorm is live. I’ll add a filter button and poll for the live leaderboard. Run Filter: worlds-fair seedstorm Polling now. Current record to beat is 5.7493835784. Ask ARIA Tsweeney's workspace Personal workspace Runs 248 Search runs 1 filter NAME NOTES [illegible] VAL_B Hypothesis: Width7... worlds-fa... 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity... 6.03687 Hypothesis: Width5... width51... 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor... 6.18444 Hypothesis: The lo... 1-23 of 23 Y-axis ticks: 6, 6.5, 7, 7.5, 8, 8.5, 9, 9.5, 10, 10.5, 11, 11.5 Interface text: Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:35:45 wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products aria-autoresearch-01 Projects wandb_agent_default_pr... Workspace Tim Sweeney wandb Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts ### Analyzing top runs ARIA examines project metrics after a request to summarize the highest-performing runs. Tim Sweeney presents this as a way for a new team member or someone returning from time off to catch up on the research. [6:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=361s) [An assistant analyzes project metrics after a request to summarize the highest-performing runs.](https://assets.ai.engineer/sites/aiecode2025/sha256/b268b1d3380e02664d30a030f1549bbb70075bdaca5ea9d49fe0cac497963079.webp) ARIA Can you please summarize the highest performing runs in this project? Thinking Analyzing project metrics shell grep -nE '^#{1,2}' skills/wbagent/references/RUN_OPS.md &&... shell sed -n '220,380p' skills/wbagent/references/RUN_OPS.md Thinking Summarizing run metrics shell python skills/wbagent/scripts/wandb_run_ops.py inspect-sch... shell python skills/wbagent/scripts/wandb_run_ops.py count-runs ... Ask ARIA Interface text: Tsweeney's workspace Personal workspace Runs 248 Search runs 1 filter NAME NOTES ITEF VAL_B Hypothesis: Width7... worlds-f... 5.83786 Hypothesis: Width6... context1 5.88779 Hypothesis: Width6... capacity... 6.03687 Hypothesis: Width5... width51... 6.0273 Hypothesis: Width5... maxdata 6.08972 Hypothesis: Test w... long-hor 6.18444 Hypothesis: The lo... 1-23 of 23 Chart axis ticks: 11.5, 11, 10.5, 10, 9.5, 9, 8.5, 8, 7.5, 7, 6.5, 6 New chat Quickly give a brief summary of th... wb-agent-team/wandb-agent-p... Can you please summarize the hig... Hello ARIA, you are on stage at AIE... Quickly give a brief summary of th... wb-agent-team/wandb-agent-p... Can you please summarize the hig... Hello ARIA, you are on stage at AIE... Yesterday Quickly give a brief summary of th... wb-agent-team/wandb-agent-p... hello there! introduce yourself to A... Quickly give a brief summary of th... wb-agent-team/wandb-agent-p... Project Workspace Runs Jobs Automat. Sweeps Reports Artifacts aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products ### A new best run family ARIA identifies four rope + SwiGLU + batch 128 frontier runs as the top four by validation loss. The best scores 4.391, about 0.14 val/loss better than the previous science-plan best. [6:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=387s) [The assistant identifies a run family with improved validation loss.](https://assets.ai.engineer/sites/aiecode2025/sha256/520ab91617689606c6a7e4e48acc4454eccac03e1159312ed47955d372579f6f.webp) [illegible] sampled val/loss history. The strongest hook is that your recent rope + SwiGLU + batch 128 frontier runs look like a clear step-change rather than a small incremental win. Hooks worth investigating 1. A new best family has emerged. The 4 rope-swiglu-b128-frontier runs are the top 4 by val/loss: best is RSB03_rope_swiglu_48x_b128_1250 at 4.391 val/loss / 6.334 val_bpb. The previous science-plan best, SP16_rope_swiglu_48x_b128, is 4.529, so the frontier set improved by about 0.14 val/loss. 2. The improvement is probably not just one lucky seed. Interface text: Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:36:25 wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products aria-autoresearch-01 Projects wandb_agent_default_pr... Workspace Tim Sweeney wandb Tsweeney's workspace Project Workspace Runs Jobs Automat. Sweeps Reports Runs 248 Search runs 1 filter NAME NOTES ITEF Hypothesis: Width7... world... Hypothesis: Width6... conte... Hypothesis: Width6... capac... Hypothesis: Width5... width... Hypothesis: Width 1-23 of [illegible] ARIA New chat Hey - what can you do? 4. AR - Workspace Builder can you please summarize the last ... AR - Quickstart Empty AR - Minimal autoresearch-demo-01 3. AR - Summary Report AR - Outlier Plot 2. AR - Find patterns /autoresearch-quickstart Ask ARIA ### Consistency and batch size ARIA finds four closely clustered frontier losses and a batch-size signal: median val/loss is 4.681 at batch size 128 versus 5.854 at 16. The report cautions that later experiment waves confound the comparison. [6:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=395s) [The report compares validation losses across runs and batch sizes.](https://assets.ai.engineer/sites/aiecode2025/sha256/d18d81135c22cc2d3c8de17656638296c4ec5fa324322c590e4124a6eee85f1e.webp) by about 0.14 val/loss. 2. The improvement is probably not just one lucky seed. The frontier quartet is tightly clustered: 4.391, 4.431, 4.435, 4.478. That consistency is more interesting than the single best run. 3. Batch size looks like a major lever. Across runs with val/loss, the best batch_size=128 run is 4.391 and median is 4.681; batch_size=64 best is 4.591 / median 4.969; batch_size=16 median is much worse at 5.854. This is confounded with later experiment waves, but it is a strong signal. Interface text: wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb Tsweeney's workspace ARIA Project Workspace Runs Jobs Automat. Sweeps Reports Runs 248 Search runs 1 filter NAME NOTES Hypothesis: Very sli... Hypothesis: Width7... Hypothesis: Width6... Hypothesis: Width6... Hypothesis: Width New chat Hey - what can you do? 4. AR - Workspace Builder can you please summarize the last ... AR - Quickstart Empty AR - Minimal autoresearch-demo-01 3. AR - Summary Report AR - Outlier Plot 2. AR - Find patterns /autoresearch-quickstart Ask ARIA ### Architecture patterns ARIA flags RoPE and SwiGLU as a promising recipe. SwiGLU runs have a median val/loss of 4.503 across 8 runs, versus 5.505 across 123 GELU runs; RoPE runs have a median of 4.478. [6:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=399s) [The report highlights RoPE and SwiGLU in high-performing runs.](https://assets.ai.engineer/sites/aiecode2025/sha256/dc727db533ebf4384fe1a25cf3f7d001d10bdf63326180c85b5083fe02fd7184.webp) Across runs with val/loss, the best batch_size=128 run is 4.391 and median is 4.681; batch_size=64 best is 4.591 / median 4.969; batch_size=16 median is much worse at 5.854. This is confounded with later experiment waves, but it is a strong signal. 4. Rope + SwiGLU appears repeatedly in the winning recipe. activation=swiglu has best 4.391 / median 4.503 across 8 runs, while gelu has best 4.697 / median 5.505 across 123 runs. pos_type=rope similarly has best 4.391 / median 4.478. This looks like a high-value architecture hook. 5. The science-plan phase already identified the recipe before the frontier phase amplified it. Interface text: wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb Tsweeney's workspace Project Workspace Runs 248 Search runs 1 filter NAME NOTES ITER Hypothesis: Very sli... Hypothesis: Width7... Hypothesis: Width6... Hypothesis: Width6... Hypothesis: Width Jobs Automat. Sweeps Reports ARIA New chat Hey - what can you do? 4. AR - Workspace Builder can you please summarize the last ... AR - Quickstart Empty AR - Minimal autoresearch-demo-01 3. AR - Summary Report AR - Outlier Plot 2. AR - Find patterns /autoresearch-quickstart Ask ARIA ### ARIA report evidence Sweeney asks ARIA to create a Weights & Biases report. The response lists loss curves, comparison charts, scatter plots, and a panel for training and validation curves and likely next ablations. [6:54](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=414s) [ARIA lists charts and panels included in a report.](https://assets.ai.engineer/sites/aiecode2025/sha256/33de3fb16dc9d753fa123345f83d4bfa4ac3a4783aed8d254bea51e7c995e292.webp) Visual evidence included in the report: • Validation-loss curves for the top runs. • Final validation-loss and best-bpb bar charts. • Science-plan architecture comparison charts. • Scatter plots for model size, batch size, and steps versus validation loss. • A frontier-focused panel section showing training/validation curves and the likely next ablations. Interface text: Chrome File Edit View History Bookmarks Profiles Tab Window Help Wed Jul 1 13:36:50 wandb.ai/aria-autoresearch-01/wandb_agent_default_project/workspace?nw=nwusertsweeney D: ARIA AIEng D: Weave - Prod Logs Metrics To Review My Stuff Monitoring My Work Preso Administration Dev Tools Products aria-autoresearch-01 > Projects > wandb_agent_default_pr... > Workspace Tim Sweeney wandb Tsweeney's workspace Project Workspace Runs Jobs Automat. Sweeps Reports Runs 248 Search runs 1 filter NAME NOTES ITER Hypothesis: Very sli... Hypothesis: Width7... Hypothesis: Width6... Hypothesis: Width6... Hypothesis: Width 1-24 of [illegible] ARIA New chat Hey - what can you do? 4. AR - Workspace Builder can you please summarize the last ... AR - Quickstart Empty AR - Minimal autoresearch-demo-01 3. AR - Summary Report AR - Outlier Plot 2. AR - Find patterns /autoresearch-quickstart Ask ARIA ## Full transcript [0:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1s) [music] [0:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=13s) Okay. Hello everyone and thank you for [0:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=15s) attending this session. My name is Tim [0:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=17s) Sweeney, a principal engineer at Weights [0:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=20s) and Biases and Coreweave. And for the [0:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=22s) next 20 minutes, we're going to talk [0:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=23s) about Arya, our new AI research and [0:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=26s) iteration agent. Let's go ahead and get [0:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=28s) started. [0:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=29s) So, uh, first off, just by way of making [0:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=32s) some noise, some clapping, uh, who here, [0:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=35s) um, identifies as an ML researcher? [0:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=37s) You're someone that trains models, [0:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=38s) trains the brain? [0:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=41s) I heard one. Wow. Okay. Great work. [0:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=43s) Great work. Uh, what about who here is [0:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=45s) the applied engineer, the namesake of [0:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=47s) this conference? Who here actually [0:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=48s) builds the bots? [0:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=50s) >> Okay, good. Expected much more. And who [0:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=53s) here is in AI management? You are [0:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=55s) helping fund this compute. [0:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=58s) Okay. Okay. Nice. From the back. Lovely. [1:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=61s) Um, well, now that I know a little bit [1:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=62s) about you, just a little bit about me. [1:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=64s) Uh, again, my name is Tim. I have a [1:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=66s) masters in machine learning, uh, and [1:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=67s) reinforcement learning from Georgia [1:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=69s) Tech. So, I've been that, uh, researcher [1:11](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=71s) currently building Weights and Biases [1:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=73s) agent, Arya. So, identify as that [1:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=75s) applied engineer. And in a previous life [1:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=78s) was the PM of Twitter's ML stack. So, I [1:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=81s) hope you hopefully can connect with you [1:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=82s) middle management as well. [laughter] [1:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=84s) >> [gasps] [1:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=84s) >> Um today's agenda is kind of broken into [1:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=87s) three sections and hopefully each of you [1:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=89s) personas walk away with something [1:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=90s) valuable. So first we're going to learn [1:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=92s) about Arya itself and how it can [1:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=94s) supercharge your AI and ML workflows. [1:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=97s) We're going to dive into auto research [1:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=98s) and see that live in a live demo in just [1:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=100s) a moment. [1:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=102s) Then we're going to pull back the [1:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=103s) curtain and learn how we use weights and [1:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=104s) biases and uh coreweave to actually [1:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=107s) build Arya because a lot of you in the [1:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=109s) audience are building agents yourself [1:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=110s) and we believe a lot of these components [1:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=112s) can help you in your endeavors. And then [1:54](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=114s) towards the end we'll just take a step [1:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=116s) back and identify a few key tips and [1:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=118s) tricks for making sure that you're able [1:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=119s) to productionize your systems [2:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=120s) effectively. [2:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=122s) For those of you who might not be [2:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=123s) familiar, Weights and Biases is the [2:05](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=125s) world's leading AI development platform. [2:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=127s) We've been in business now for nine [2:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=129s) years and have happily joined the core [2:11](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=131s) family about a year ago. Uh we have a [2:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=133s) number of products in our suite but are [2:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=135s) really known for our models training [2:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=137s) inference and weave stack which really [2:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=139s) helps collect data uh about the AI [2:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=141s) development and machine learning [2:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=142s) workflows and makes that information [2:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=144s) actionable and uh enables users to make [2:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=147s) the best decisions about what to do [2:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=148s) next. [2:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=150s) So without further ado, let's go ahead [2:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=152s) and dive into Arya, our agent. Uh we'll [2:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=154s) show a demo and then we'll get back to [2:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=155s) some slides. [2:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=160s) Okay, beautiful. [2:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=162s) Let's make this a bit bigger. Holler at [2:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=164s) me if you need it to be bigger. So, uh, [2:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=166s) what you're looking at here is a weights [2:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=168s) and biases workspace. For you, for [2:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=170s) anybody that isn't familiar, on the [2:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=171s) lefth hand side, I actually see a list [2:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=173s) of a bunch of different experiments. In [2:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=175s) this particular project, I have over 200 [2:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=177s) training jobs. And on the right hand [3:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=180s) side, I see a scatter plot of, in this [3:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=182s) case, declining metrics, which is good. [3:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=184s) means our loss is going down over time. [3:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=186s) And this view would be very familiar for [3:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=188s) anyone that uses our tool. Now, to [3:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=190s) ground this, we're actually uh uh using [3:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=193s) the Carpathy Auto Research project, [3:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=195s) which I'm sure many of you are familiar [3:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=196s) with, but if you're not, it's just a [3:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=198s) very simple project that trains an LLM, [3:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=200s) and it's a great foundation for auto [3:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=203s) research type demonstrations because [3:25](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=205s) it's a very simple codebase and allows [3:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=207s) us to improve iteratively over time. So, [3:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=210s) let's jump back to the project and open [3:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=211s) up Arya by clicking this blue button in [3:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=213s) the upper right. When I click this [3:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=215s) button, I'm uh presented with the [3:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=217s) familiar chat interface with, you know, [3:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=219s) how can I help you today, a few call to [3:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=220s) actions, and you know, I can add [3:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=223s) different context in my project or maybe [3:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=225s) add images, etc. Um, everyone here is [3:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=228s) agent builders, so I don't need to bore [3:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=229s) you with the details of what an agent [3:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=231s) interface looks like. But let's go ahead [3:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=233s) and just, you know, enter in a basic [3:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=235s) intro here. Let's say, "Hello, Arya. [3:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=238s) You're on stage at AI World's Fair 2026. [4:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=240s) Please introduce yourself." So, it's [4:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=242s) going to go ahead and chug along and [4:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=243s) hopefully emit some sort of nice emoji. [4:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=246s) Yay. He I'm Arya. I'm talking to the [4:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=248s) audience. Great. But now, let's dive [4:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=250s) into the meat of why you came here. So, [4:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=253s) I'm going to open up this chat here. And [4:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=255s) this is a longunning chat where I've [4:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=257s) been running again over 200 experiments [4:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=259s) using the auto research loop. Um, it [4:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=262s) helped me download the code, set up my [4:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=264s) launch jobs, set up my GPUs, and is able [4:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=266s) to autonomously iterate on the code [4:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=268s) itself and the hyperparameters. [4:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=270s) We'll take a look at what it's doing in [4:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=271s) a moment, but while we're doing this, [4:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=273s) I'm going to kick off a live iteration [4:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=275s) right here. So, what I'm going to say is [4:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=277s) please conduct another batch of [4:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=279s) experiments. You are on stage at the AI [4:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=281s) Engineer Worlds Fair 2026 and we're [4:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=284s) hoping to find the best model live. I [4:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=286s) believe in you. Uh, because we know we [4:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=288s) have to encourage our models. Um, so [4:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=290s) it's been doing this for a while. What [4:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=291s) it what it's doing here is it's saying, [4:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=293s) "Okay, great. Um, I don't want to make a [4:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=295s) big architecture swing. That feels a [4:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=297s) little bit too risky." So, it's probably [4:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=298s) going to go for uh some modifications to [5:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=301s) the hyperparameters. And then it's [5:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=303s) kicking off a shell call here that is [5:05](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=305s) actually um executing that uh executing [5:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=308s) that experimentation loop. And we're [5:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=310s) going to check in on this periodically [5:11](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=311s) throughout this presentation. [5:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=313s) But I want to help explain what's going [5:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=315s) on behind the scenes. So behind the [5:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=317s) scenes I have set up a weights and [5:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=319s) biases launch queue. Launch is our our [5:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=321s) product that allows you to connect your [5:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=323s) compute clusters and allows humans and [5:25](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=325s) agents to launch longunning [5:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=327s) experimentation jobs particularly by [5:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=329s) leveraging GPUs. [5:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=331s) Here I'm looking at a uh a terminal [5:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=333s) output of my Kubernetes cluster where [5:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=335s) we're actually seeing live execution of [5:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=337s) experiments happening. So this is [5:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=339s) happening live right here. This is not a [5:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=340s) fake demo. Um great. And if we jump [5:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=344s) back, we see that at this point it [5:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=346s) started the cues and now it is simply [5:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=348s) polling and waiting for our work to be [5:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=350s) complete. So we'll jump back to that in [5:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=352s) a in a moment. But before but let's dive [5:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=355s) into a few other examples. So uh [5:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=357s) something else that is interesting you [5:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=359s) can do is maybe you might want to ask it [6:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=362s) something like please summarize the [6:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=363s) highest performing runs in this project. [6:05](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=365s) This use case would be something like [6:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=366s) maybe a new user come or a new uh team [6:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=368s) member is joining your project and want [6:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=370s) to understand the research. Um or maybe [6:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=372s) you've uh someone's been doing some work [6:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=374s) while you were on PTO and you want to [6:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=376s) get caught up. We'll see what this comes [6:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=377s) up with in a moment. Some other [6:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=379s) pre-anned uh examples are finding [6:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=381s) patterns in your project. So here we can [6:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=384s) see that I asked it, hey, can you find [6:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=386s) some patterns in this research? And we [6:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=388s) see that um it identified that a new [6:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=390s) family of models emerged as the as the [6:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=393s) auto uh auto research was happening. Uh [6:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=395s) it identified that batch size seems to [6:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=397s) be a really high high uh lever uh [6:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=400s) parameter. It identified an [6:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=402s) architectural recipe that seemed to be [6:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=404s) quite promising and a number of other [6:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=406s) insights that would have taken me hours [6:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=408s) or days to discover on my own. And Arya [6:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=410s) is able to do it right for me directly [6:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=412s) in the interface that I already live. [6:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=415s) Not only is it able to emit text based [6:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=417s) uh textbased outputs, but it also deeply [7:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=420s) integrates with a number of weights and [7:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=421s) biases visualization utilities. So here [7:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=424s) I've actually asked it to emit a weights [7:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=426s) and biases report which for those who [7:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=428s) aren't familiar is essentially a [7:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=429s) markdown file on steroids. It's got uh [7:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=432s) embedded embedded plots, charts and and [7:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=434s) and graphics. And so here uh you know [7:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=436s) it's talked about the thesis of the [7:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=438s) project. It's it's emitted a number of [7:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=440s) of data panels. And uh I actually think [7:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=443s) it's quite interesting. It used um one [7:25](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=445s) of our more esoteric panels, the uh [7:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=447s) parameter importance chart to uh tell me [7:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=450s) the correlation of various different [7:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=452s) parameters within this uh within this [7:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=453s) training job. [7:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=455s) Uh in addition to uh reports, it's also [7:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=458s) great at working with workspaces. So if [7:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=460s) you're a weights and biases user, uh you [7:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=463s) spend a lot of your time uh designing [7:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=465s) and working with workspaces. Well, Arya [7:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=468s) is actually customtuned and prompted to [7:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=470s) really understand how to build [7:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=471s) workspaces, build plots, and complement [7:54](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=474s) that that data analytics with real live [7:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=476s) graphics using the built-in proprietary [7:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=478s) charts that weights and biases users [8:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=480s) know and love. Um, so with that, let's [8:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=483s) go ahead and check back on some of our [8:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=484s) our prompts. We can see that the please [8:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=486s) summarize this project prompt is cooking [8:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=488s) away. It's querying weights and biases. [8:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=490s) It's applying patches. It's writing its [8:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=492s) own code. So, we'll come back and check [8:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=493s) on that in a moment. and our longunning [8:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=495s) training job is uh still pulling for the [8:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=498s) results. We can see that we're cooking [8:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=500s) away on our GPUs. So, we're we're frying [8:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=502s) some GPUs [music] and doing some data [8:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=504s) science all live. And while that's [8:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=506s) cooking, let's go ahead and jump back to [8:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=507s) the presentation. We'll come back in a [8:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=508s) moment. [8:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=510s) Uh [8:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=513s) oh, no, we're not looking at a [8:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=514s) dictionary. We're looking at a PO. [8:36](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=516s) Great. Uh okay, so quick recap here. [8:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=518s) What did Arya show? What did we show in [8:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=520s) these last five minutes? First, we show [8:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=522s) that uh Arya can serve as your data [8:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=524s) science companion right inside of [8:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=526s) Weights and Biases, helping you discover [8:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=528s) insights that you wouldn't you wouldn't [8:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=529s) be able to discover as your experiments [8:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=531s) and as your team size grows. [8:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=535s) Next, we address the problem of [8:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=537s) complicated reporting and complicated [8:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=538s) plotting. Weights and biases users are [9:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=540s) are really want to turn their insights [9:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=542s) into visual communication tools. They [9:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=544s) want to communicate with their peers and [9:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=546s) their colleagues. So Arya's built from [9:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=548s) the ground up to understand those [9:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=549s) primitives and help co-pilot and drive [9:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=552s) right along right alongside in the UI [9:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=555s) and announcing now today for the first [9:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=557s) time we are releasing Arya on our iOS [9:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=559s) device or on our iOS app. So uh uh Arya [9:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=563s) released on Monday and our iOS app now [9:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=566s) has Arya built in. So if you're [9:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=567s) conducting hyperparameter tuning jobs, [9:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=569s) if you're training models, or if you're [9:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=571s) just researching within the weights and [9:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=572s) biases ecosystem, you can go touched [9:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=574s) grass at Yerba Buena uh gardens and [9:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=577s) steer your uh hyperparameter tuning jobs [9:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=579s) all from your mobile device. And what is [9:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=581s) this all building up to? This is [9:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=583s) building up to an a fully automated [9:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=585s) endto-end research platform where we're [9:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=587s) not seeking to replace uh RL [9:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=589s) researchers, but complement your [9:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=590s) workflows. Arya's great at orchestrating [9:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=593s) jobs, understanding GPU workloads, [9:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=595s) responding to events within the within [9:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=597s) the Wandi ecosystem, and listening to [9:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=599s) researchers, uh, uh, looking up archive [10:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=601s) papers, and collaborating on hypothesis. [10:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=604s) So, we can let Arya drive the mechanics [10:05](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=605s) that you don't want to deal with while [10:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=607s) you focus on the new ideas, new [10:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=608s) architectures, and new parameters that [10:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=610s) you wanted to try. [10:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=612s) Um, great. So, that's Arya in a [10:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=614s) nutshell. We're really hoping that you [10:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=616s) give it a shot. And uh we'll jump back [10:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=618s) to the auto research at the end and see [10:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=620s) if we got a new best record. But before [10:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=622s) we do that, let's talk about how we use [10:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=624s) weights and biases and coreweave to [10:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=626s) actually build Arya. So now speaking to [10:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=628s) a lot of the the AI agent builders in [10:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=630s) the room, here's a quick architecture on [10:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=632s) the lefth hand side. You see that we [10:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=634s) have a web client, iOS client that [10:36](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=636s) communicates with our API server that [10:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=637s) then dumps data into our turn database [10:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=639s) and is worked on by our harness, our our [10:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=641s) worker harness. This is sort of [10:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=643s) archetypical of probably what most of [10:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=645s) you are all building in the room and is [10:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=647s) exactly what we have on our back end. [10:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=649s) But that harness worker is a magic is a [10:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=651s) is a magic box and it connects to a [10:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=653s) number of important utilities. First is [10:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=655s) a sandbox where it can execute arbitrary [10:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=657s) shell calls uh do do Python data science [11:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=660s) etc. And we invite you to try coreweave [11:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=662s) weights and biases sandbox to fit into [11:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=664s) your architecture. [11:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=666s) Next up you need an LLM provider of [11:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=668s) course and so if you're maybe using GLM [11:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=670s) 5.2 two or one of your fine-tuned [11:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=672s) models. We invite you to use uh weights [11:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=674s) and biases inference and connect that to [11:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=676s) your worker as well. [11:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=678s) If you're like us, you need to run [11:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=680s) longunning workloads outside of the main [11:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=682s) loop of the agent where you're actually [11:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=684s) training for day for sometimes days at a [11:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=686s) time. Weights and biases launch can [11:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=688s) actually help facilitate that and [11:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=689s) coreweave GPUs can help make that [11:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=691s) compute even better. [11:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=693s) And then lastly, and really most [11:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=695s) importantly, we need an observability [11:36](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=696s) layer. It's critical that your agents [11:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=698s) are able to log out their what's going [11:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=700s) on with their sessions, their turns, [11:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=702s) their tool calls, any errors they're hap [11:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=703s) that that's happening, etc. Uh we have a [11:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=706s) product called Weights and Biases Weave [11:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=708s) that we log 100% of our traces to where [11:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=710s) us and our team can learn from. And [11:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=712s) that's where we move from production to [11:54](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=714s) offline where our team is able to use [11:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=716s) Weights and Biases Weave to drive [11:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=718s) insights and identify behaviors, [12:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=720s) implement tasks with tasks which are [12:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=722s) essentially unit tests for your models [12:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=724s) and evaluate those models in a loop. [12:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=727s) We have a model repository which you [12:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=728s) might choose to use weights and biases [12:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=730s) artifacts to store your agents or models [12:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=732s) and you we emit our evaluation results [12:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=735s) to weave where we have a common [12:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=736s) dashboard that we can make go no-go [12:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=738s) decisions on various prompt changes or [12:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=740s) architectural changes that then feeds [12:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=743s) into a research loop which we call our [12:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=744s) improvement loop where we form [12:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=746s) hypotheses implement candidate agents [12:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=748s) and analyze the evals. So we have two [12:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=750s) sort of complimentary yet adversarial [12:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=752s) research loops going on going on offline [12:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=755s) feeding data from weights and biases [12:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=757s) weave ultimately to identify the best [12:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=759s) model so that we can promote that to [12:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=761s) production through our registry and [12:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=762s) close the data flywheel. So in the next [12:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=765s) just uh three seven minutes or so we'll [12:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=767s) just talk about uh weights and biases [12:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=769s) weave and show how we as a team actually [12:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=771s) use weave to facilitate this workflow [12:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=773s) and we believe this is something that [12:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=775s) you would benefit from as well all of [12:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=776s) you agent builders in the room. [12:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=779s) Yes, another demo. Great. [13:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=783s) Okay. Okay, we have new responses. So, [13:05](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=785s) it's going to be exciting when we open [13:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=786s) this up later. See if uh we've got some [13:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=789s) better metrics. Um, okay. Let me zoom [13:11](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=791s) out just a little bit here. So, here I'm [13:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=793s) looking at the agent dashboard. This is [13:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=795s) the live weights and biases agent or [13:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=798s) Arya agent dashboard uh built in weave. [13:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=801s) Man, that is a lot of uh branded [13:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=803s) buzzwords there. This is the dashboard [13:25](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=805s) that you would get if you use our tool. [13:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=806s) and uh you have a you know uh span [13:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=809s) volume, conversation volume, token [13:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=811s) tracking, etc. Think of this as like a [13:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=813s) uh a bird's eye view of your agent. For [13:36](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=816s) me, however, I really like this [13:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=818s) conversations view, which I do have [13:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=819s) pre-loaded in this tab. This [13:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=821s) conversations view is a live feed of all [13:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=824s) of the conversations that are going [13:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=825s) through Arya, but it's filtered down to [13:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=827s) just the internal employees. So, it's a [13:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=829s) little bit of a of a reduced set here. [13:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=831s) Um what I what I love is this middle [13:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=833s) spans view which gives me a visual [13:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=836s) indicator of the topology of a trace. [13:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=838s) Different colors and and shapes indicate [14:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=841s) different things that are happening [14:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=842s) within the agent. So things like tool [14:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=844s) calls, LLM calls, thinking blocks, etc. [14:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=846s) which really help me understand again [14:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=848s) the shape and topology of that [14:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=850s) particular conversation. I can of course [14:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=852s) open up one of these conversations and [14:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=854s) view our our conversation view where I [14:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=857s) can see the system prompt, the user [14:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=858s) message, shell calls, reasoning blocks, [14:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=861s) etc. This is where my research lead, [14:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=863s) myself and my PM go to add notes, add [14:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=866s) feedback, add emojis, and talk about and [14:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=868s) discover those insights and those [14:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=870s) behavioral nuances we spoke about [14:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=872s) earlier so that we can turn them into [14:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=873s) tasks. [14:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=875s) Arya's built in to the weights and [14:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=877s) biases system as well. Here you'll see a [14:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=879s) summarize button and these are sprinkled [14:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=880s) throughout the weights and biases [14:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=882s) application. I simply click summarize [14:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=884s) and we start a new chat contextualized [14:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=887s) to the thing that I'm looking at. So it [14:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=889s) it sees this and says give me a brief [14:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=891s) summary of this particular conversation. [14:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=893s) So if you if you're paying a attention [14:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=895s) closely, you'll realize that what we're [14:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=897s) doing is using Arya to analyze Arya's [15:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=900s) own conversations to then make [15:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=901s) recommendations about how to improve [15:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=902s) Arya all within the UI. [15:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=906s) Um okay, great. While that's cooking [15:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=908s) away, I want to show you the last item [15:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=910s) uh within the Weave ecosystem here, and [15:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=912s) that's signals. We've heard a lot today [15:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=914s) about the value of evals and the value [15:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=916s) of LLM judges. Weave actually offers an [15:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=918s) integrated LLM judge experience. So [15:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=921s) here, if I zoom out a little bit, you'll [15:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=923s) see that I have a user frustration [15:25](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=925s) signal, a lowquality response signal, [15:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=927s) ask user signal, etc. These are LLM [15:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=929s) judges that run live against against our [15:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=932s) live traffic. And we can see various [15:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=934s) different signals like user frustration [15:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=935s) moments or lowquality responses. These [15:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=938s) help our team identify these clusters of [15:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=940s) behavior for us to go fix in next week's [15:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=942s) iteration. Let's go ahead and do a live [15:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=944s) look and see what it says. Um this says [15:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=947s) the user explicitly states that I'm not [15:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=949s) satisfied with the loss curve. It looks [15:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=951s) bad and it apparently that indicates [15:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=953s) frustration. So here we can see an LLM [15:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=956s) judges live reasoning for why that [15:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=958s) particular flag was uh indicated. [16:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=962s) Uh let's see, four minutes left. [16:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=963s) Perfect. Um so, uh with that, I've been [16:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=966s) using the term task a lot. And so what [16:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=968s) we're do, what I've showed so far is is [16:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=970s) this live production loop where we are [16:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=972s) are are are tracing our our prod logs. [16:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=974s) We're looking at them as humans, maybe [16:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=976s) even using LLMs to complement that [16:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=977s) analysis. And what we end up doing is [16:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=980s) transforming those into tasks. Now, this [16:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=982s) gets a bit technical here, but our tasks [16:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=984s) are all described as YAML files. You can [16:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=986s) think of a task as essentially a unit [16:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=988s) test for your model. So here we say we [16:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=990s) have a an example user prompt that says [16:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=993s) check this run and that run. Both of [16:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=995s) these are giving good results. What can [16:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=997s) we learn from this? What's the [16:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=998s) difference? So this is an example of [16:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1000s) something we want Arya to be good at for [16:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1002s) all of you. And after the uh requisite [16:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1005s) metadata we see that we've defined an [16:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1007s) LLM judge. So here we've defined what [16:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1010s) correctness means in the context of that [16:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1012s) question. [16:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1013s) And we've then we've defined a second [16:55](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1015s) LLM judge that determines if the [16:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1018s) insights are actually interesting. [17:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1020s) [laughter] And then we've uh defined a [17:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1022s) third rule-based judge that says were [17:05](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1025s) you able to actually generate a result [17:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1027s) within just six tool calls meaning it [17:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1029s) got there with some degree of [17:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1030s) expediency. These are all then clustered [17:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1033s) together into we have about like 200 of [17:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1035s) these. They're all clustered together [17:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1037s) into an eval suite that runs nightly. [17:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1039s) And again we use weave to track all [17:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1041s) those evals. So here, I know it's a bit [17:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1043s) small on this screen, but what you're [17:25](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1045s) looking at is a listing of every night's [17:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1047s) eval. This is literally two nights ago, [17:29](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1049s) the evaluation for our candidate model [17:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1051s) got 73% on our production or on our eval [17:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1054s) suite against the 72% that our prod [17:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1057s) model got, which means we're definitely [17:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1058s) going to push that forward uh this [17:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1060s) Friday. Uh and we can see a kind of a a [17:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1062s) performance plot on the right. So these [17:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1064s) utilities are what you would get out of [17:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1065s) the box if you're uh if you decide to [17:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1067s) pick up weave and use this tool. Um, [17:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1070s) jumping back to the last conversation we [17:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1072s) had where it asked me where we asked, [17:54](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1074s) uh, can you please give a quick summary [17:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1076s) of this trace, we see that it actually [17:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1078s) analyzed the conversation, understood [18:00](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1080s) what the user was doing, and then [18:02](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1082s) ultimately decided that this was a [18:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1084s) pretty strong trace. Um, let's see, [18:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1087s) we've got two and a half minutes left, [18:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1088s) so let's just quickly recap here. Uh, [18:11](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1091s) first off, uh, what we use weave to do [18:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1093s) is a, collect production traffic. Super [18:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1095s) critical to collect all of your [18:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1096s) production traffic so you can learn and [18:17](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1097s) iterate. Secondly, we use it to generate [18:20](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1100s) insights both as humans as well. We we [18:22](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1102s) do it as humans. We use Arya and we use [18:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1104s) LLM judges to identify those behavioral [18:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1106s) nuances. We then enrich our tasks. We [18:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1110s) implement models and we evaluate using [18:32](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1112s) weights and biases weave as a shared [18:34](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1114s) dashboard where we can make decisions [18:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1115s) together as a team that then ultimately [18:38](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1118s) allows us to promote the best model [18:40](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1120s) forward with confidence. [18:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1123s) So speaking of confident [18:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1124s) productionization, let me speak uh [18:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1126s) briefly to the managers in the room. So [18:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1128s) a few tips for being successful here. [18:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1131s) First is um invest in agent-oriented [18:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1133s) observability. Uh I'm a bit biased. I [18:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1136s) believe that weights and biases weave is [18:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1137s) the uh observability platform of the [18:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1139s) future. Uh but pick your favorite [19:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1141s) flavor. Whatever it is, log your [19:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1143s) sessions, log your turns, log your tools [19:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1144s) and feedback. This introduces an ability [19:07](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1147s) to catch a new class of bugs in our [19:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1148s) world called behavioral bugs. Not [19:10](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1150s) exceptions, not performance, but [19:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1152s) behavioral bugs. [19:14](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1154s) Next up, tasks and evals are the new [19:16](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1156s) world of CI. You've heard a lot about [19:18](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1158s) this. If you are a software engineer, [19:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1159s) you've written unit tests your whole [19:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1161s) life. You must develop a practice where [19:23](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1163s) your researchers are sitting on the same [19:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1164s) scrum team as you developing tasks and [19:26](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1166s) you're viewing the performance metrics [19:28](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1168s) as true go no-go decisions. But in order [19:31](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1171s) to complement that, you must use humans [19:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1173s) as a necessary judge. There are [19:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1175s) behavioral nuances that LLM will not [19:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1177s) catch. You must be using your product [19:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1179s) and you must be manually reviewing these [19:41](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1181s) traces as a team at the end of the week [19:43](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1183s) on a board looking at the best and worst [19:45](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1185s) traces to understand how your model is [19:47](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1187s) performing. [19:49](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1189s) And then lastly, um just maybe one one [19:51](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1191s) more tip is to add value through context [19:53](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1193s) and tools. It can be really tempting to [19:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1196s) uh try to overengineer the harness and [19:58](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1198s) do a bunch of creative stuff around [19:59](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1199s) memory and things like this. We found [20:01](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1201s) that a a lot of lowhanging fruit can be [20:03](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1203s) ascertained through simply giving your [20:04](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1204s) agent context about your business [20:06](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1206s) domain, the underlying uh primitives [20:08](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1208s) that you have available and your [20:09](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1209s) particular uh business data. Um so with [20:12](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1212s) that, let's go ahead and check in on our [20:15](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1215s) uh our our research agent here and let's [20:19](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1219s) go ahead and toggle our workspace. And [20:21](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1221s) what we should be seeing is yes indeed a [20:24](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1224s) little dot that uh oh okay our previous [20:27](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1227s) dot which was done at lunch was 5.83. [20:30](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1230s) 831. This got 5.833. So we were right on [20:33](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1233s) the edge of having a live improvement, [20:35](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1235s) but pretty darn close. Uh so that's what [20:37](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1237s) the uh that's what the model was able to [20:39](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1239s) produce. It actually uh ran uh quite a [20:42](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1242s) few tests here. I see I'm over time, so [20:44](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1244s) I will click close pretty soon. But we [20:46](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1246s) ran 12 different experiments within that [20:48](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1248s) experiment batch and uh we'll be running [20:50](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1250s) more all night. So please try out Arya, [20:52](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1252s) scan the QR codes, check out the docs. [20:54](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1254s) Uh we really love to see what you do [20:56](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1256s) with it and um looking forward to [20:57](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1257s) serving you. Thank you very much. [21:13](https://www.youtube.com/watch?v=hd7TOvmyAxU&t=1273s) >> [music]