36,915 results · Most relevant
When the downstream table needs to handle a wider type, update the downstream table first so its column can handle the wider type, and only then merge and deploy the PR. This ordering avoids incidents from type-widening changes.
When a column gets re-tagged as sensitive, search the code base for anyone still reading it from the copy that's about to lose it. This became a new rule after the re-tag miss.
Robinhood built a layered PR review system where deterministic scripts, an AI agent, and humans each handle different review responsibilities. About 60% of data PRs now merge without human review, but protected decisions still reach a person by design — the goal is less review work, not removing accountability.
Zach Lloyd says the verification step is very important for certain types of apps like UIs but less critical for others. For UI apps, use computer use so the computer actually uses the code the agent produced and produces videos and screenshots — skipping this misses problems it would catch. Also still use CI/CD in the factory.
The harness can be thought of as a form of domain specificity: when deploying a robot in a new environment (e.g., a wet lab picking up tests), a lot can be learned in context. There is a relationship to meta-learning literature: a bigger model programs a smaller model to do things, and the smaller models learn in context wrapped inside the harness. Smaller models inside the harness can be made smaller and run faster as long as the bigger model can do the domain specialization of the harness well.
almost seems like; can be made maybe smaller
Vlad Tenev describes the Trump accounts program, which gives every child born in January 2025 onward (for the next three years) $1,000 seeded by Treasury, with the idea that every child from birth can own a chunk of the US economy through a low-cost diversified S&P 500 ETF. Robinhood was selected as the sole initial trustee and broker for the program, working with BNY (America's oldest and most storied bank), Treasury, and the IRS. Robinhood is powering the infrastructure and building the app, which they see as their way of serving their country.
When a workflow element does tool calls, Ramakrishnan Lokanathan says the eval objective should be: what is your tool call ability and the tool call accuracy at that point. The eval has to be designed for tool calling and not for a different process separately.
The creator advises incorporating touching into the date and having moments where you look deep into each other's eyes, presenting these as key elements that build attraction and intimacy.
Zach Lloyd argues that storing all agent traces and artifacts — recording everything the team's coding agents do — is what makes coding agent activity the raw material for improving the factory; without it, you lose the raw material needed to improve over time.
Supabase is announcing Supabase Compute, a long-running serverless compute offering that runs in the same private network as the user's database for co-location and low latency, is a full compute instance, can run agents, supports long-running workloads like chunking and embedding, and allows direct Postgres connections from compute.
Kevin Hou announces that the agent manager is now a standalone app decoupled from the IDE in Antigravity 2.0, and describes its core purpose of managing many agents in parallel with the human in the driver's seat.
To leverage new models same-day, Optiver benchmarks them on golden sets (requiring a mature evaluation framework), partnering with key customer trading teams to collect and onboard their benchmarks so new models can be run immediately and results shared company-wide. A new model must be made available in the gateway with the right APIs, interfaces, and cost management; harnesses often need updating, which Optiver manages and delivers; and after benchmarking they keep the communication line open — what they looked into, how well it works, how to use it, and their recommendation.
Optiver's principle: if a new frontier model is made available publicly, they want to leverage it in production flows the same day, because there is immense value in that if it is valuable.
When scoring behavior, Irina measures source retention among agents who know the fact (provenance), checks whether uncertainty survived propagation ('might leave' staying 'might leave'), and checks whether agents acted on what they actually knew.
hedged
Irina maps scoring failures to policy changes: if source attribution disappears, preserve source in memory writes and summaries; if rumors harden into facts, store confidence, mark first-hand versus second-hand, and require hedging when retelling uncertain claims; if public facts stay local, classify useful public facts differently and make agents proactively share important source evidence.
hedged
Irina gives three concrete controlled-scenario designs: (1) public fact diffusion — Agent A learns the bakery will close tomorrow and the test checks whether the right agents learn it, remember who said what, and change plans; (2) rumor uncertainty — Agent A hears Agent C might leave the village and the test checks whether 'might leave' hardens into 'is leaving'; (3) replanning — the group has a plan, one agent learns the route is blocked, and the test checks whether agents update and communicate this.
hedged
You can ask ARIA to summarize the highest performing runs in a project, useful when a new team member joins or when catching up after PTO. You can also ask ARIA to find patterns in your research project; it identified that a new family of models emerged, that batch size seems to be a really high lever parameter, and an architectural recipe that seemed to be quite promising.
hedged
An agent that retries a request without checking whether the first operation committed can create duplicate effects (e.g., two tickets when only one was intended).
When lexical search fails to rank relevance, the model ends up reading every document, and the more documents there are, the more a basic query can cost $10-20 in tokens.
Rob Snyder recommends having two slides in a pull-validation conversation: slide one restates what you heard about their project, and slide two is a conceptual overview of how you can help. This keeps the conversation anchored on the prospect's project rather than a product pitch.
Choose sources → to get a library of your own and a daily brief.