35,595 results · Most relevant
Atlassian turned the popular whiteboard prompts into dedicated features: grouping sticky notes into common themes, and turning a whiteboard into backlog tickets (users can now select specific cards and choose which to add to their Jira backlog). Both were built for humans and then exposed as tools so agents can use them too — a pattern the team says it 'accidentally fell into.'
Roughly 1,000 concurrent environments were run to generate trajectories, and this all cost under $500 (roughly $420 compute, $40 for the Judge API) on H100.
Bad scope occurs when memories are not appropriately scoped to the right context — like logging into your ChatGPT account and getting someone else's message history. Improper memory scoping can cause serious data leaks. Memory scoping needs to be applied on both write and on read.
In the demo, the agent's plan is given to a specialized tool called plan and compile, which compiles it into a very deterministic workflow. The compiled output is a fully runnable workflow that gets executed, relying on Conductor as the workflow execution engine.
Pat Gelsinger argues extreme specialization in compute architecture is a mistake because workloads will continue to migrate significantly over the next two to four years. Compute fleets are becoming more granular across the workload — specialized prefill, decode, midfill, speculate-and-verify tiers — and such extreme heterogeneity has never been sustainable historically. Reasoning models shift demand back toward CPU-like compute, changing which portion of compute capacity is predominant.
Pat Gelsinger says chip development still takes a year and a half to get into racks at scale, which remains the first bottleneck regardless of AI design tools. If it takes that long, the designer's understanding of AI workloads is no longer applicable to the chip they designed — Graphcore wasn't a bad design, but the world moved on. He wants to compress prototyping to a month or two and design to three months because slow cycles make designs obsolete.
The complete engineering analysis found cycles of dependencies, risky long-running migrations, single points of failure, and blind spots in telemetry — issues the speaker frames as incidents waiting to happen.
The Wow CLI is a toolkit that sets up Git, creates a repo in the enterprise GitHub account from a predefined template (a blank Node.js app), and clones an example application for Claude Code or Codex to build. The 'wow app create' command names the app, scans the machine, and checks whether Git is installed, GitHub is logged in, and Doppler (secrets management) is present — pre-configuring the machine so engineers and non-engineers alike hit no snags. After scaffolding, you can pop into Claude and ask what to put in the app (e.g., 'I'm doing a demo for AIE, what can we put in this app that would be cool?').
described
The speakers report that from the beginning, their hypothesis has been that one of the big bottlenecks is automatic characterization, and they focused on it — closing the loop with simulations — because that is how you try many things, make informed decisions, and decide what to do the next day.
described
The guests describe building up a large basket of experimental data and 'timestamping the state of the world' — recording what the experimental evidence was as of a given date. From such a timestamped body of evidence, they can construct RL environments asking what the next choice the scientists made was, or what the outcome of an experiment was. Having lineage to the whole process enables constructing new types of environments that are not plausible elsewhere.
described as the lab's data strategy
The guests explain that they think of their AI program as constructing agents around the set of data they have produced. RL environments can be built for individual pieces of the loop — for example, rewarding identification of phases present and penalizing spurious or chemically implausible phases in phase identification — and stitching these together yields the end-to-end discovery loop.
described as the lab's approach
Kyle explains that when she's leaving soon — e.g., last night of vacation with dinner plans — the best close is proposing a tentative meetup that fits her archetype and reality, such as grabbing a drink or meeting at a venue, rather than forcing a formal date.
When soccer clubs contact him about high-speed injuries, he observes players who are tilted forward the entire sprint, or whose sprint is fine until fatigue causes them to tilt forward; soon after, they grab their hip or thigh and are out with injury.
anecdotal observation from his consulting experience
The 'npx skills' command is the most popular way to install skills — and it's the only command that installs not just a skill but also free prompt injections, a wild script requiring root permissions, and someone else's API key all in one run. RunLayer scans thousands of skills every day and there is really weird stuff inside them. Scan and vet third-party skills before installing them.
Joe Rogan proposes what he sees as the only way to get everyone on your side: borders are closed, and you find out which illegal immigrants have committed felonies. If an illegal immigrant has committed X amount of felonies, they have to leave the country — if you've been robbing people for the last 10 years and you're an illegal, you have to leave the country, 'that makes sense.' But if you've been here for 25 years, have a family, kids go to school here, speak the language, and are a contributing member of the community, deporting them is crazy.
Rogan's endorsed proposal, clearly attributed.
Joe Rogan argues that when people have been here for 20 years, mass deportation is crazy — 'come on. That's crazy.' He says if immigrants have been productive members of society for 20 years — no criminal record, worked the entire time, paid taxes — find them a pathway to citizenship. He also notes that shipping lower-income and middle-income people who are kind of getting by to a country where they've never lived is crazy to ask, and that deporting long-term contributing illegal residents who have families and kids in school shows no heart — you won't get any reasonable people to go along with it.
Rogan's endorsed position, clearly attributed.
Fractile's key architectural bet is combining high-capacity, lower-cost DRAM (GPU/TPU-style scalability) with the speed advantages of SRAM-based chips (Grok/Cerebras-style); partnering directly with memory vendors — not just logic foundries — enabled a platform ramping in the second half of next year.
To enter mortgage servicing from scratch you need three things: get licensed, turn decades of regulation into code, and convince asset managers and lenders to hand over the regulated asset to a new company.
described
Joe Liemandt estimates that a motivated student can catch up one grade level in 20-30 hours, and that a student three years behind can catch up with an hour a day of homework in about 90 days; teaching years ahead of capability takes many more hours.
For complex document search in insurance, finance, healthcare (uncategorized or semi-labeled data), use a five-tool framework: compass-direction retrieval, listing files, traversing by metadata, grepping for keywords, and reading document content.
Choose sources → to get a library of your own and a daily brief.