Scrollback

I know people in sales can relate to this one...

Nothing’s Off The Table with Louis Ruggiero · @nothingsoffpod

It’s a beautiful day to learn about AI inference engineering

Madison Kanna · @Madisonkanna

i’m increasingly convinced that the best agent evals will come from mining real agent failure traces. my view is that every failed trace contains a potential eval but not in its raw form. raw traces are messy, long and too specific. the research problem is to distill them into clean reproducible tests. the pipeline i’m interested in is (which i'm currently working on): failure trace → failure attribution → earliest divergence point → minimal reproducible state → targeted eval → regression suite this turns trace data from passive observability into an active improvement loop. like can we extract the exact decision point where the agent should have behaved differently? and can we convert that into an eval that catches the same failure class in the future? i guess this matters because most agent failures are trajectory-level failures and not just output-level failures. personally i think this is much more realistic than relying only on hand-written benchmarks (imo they should look more like failure memory systems). hand-written evals encode what we think agents will fail on. traces encode what agents actually failed on. also once you have the mechanism, you can mutate the trace into variants. that is basically fuzzing for agents.

λux · @novasarc01

On quality: My first day at @baseten, the very first person I met was @philipkiely on the AI Education team. He told me he's working on a textbook. An absurd thing to say on the surface. I asked for an early pdf version that I started to read on the plane ride home. It was reviewed for accuracy, comprehensiveness, and clarity by dozens of people across multiple disciplines and hand-edited. The book became a useful reference for me and deeply accessible to someone new to the space and something I still return to often. A few months later, the book was complete. Printed by an old school shop in Europe, the cover, charts, and diagrams meticulously crafted pixel-by-pixel with @lucas_dehaas's hands. Sourcing the printer and getting the proofs just right added many weeks to the launch beyond the original deadline. Then, it was time for the launch. The team wanted to get the books in the hands of ML researchers and the broader tech community. Instead of shipping it via USPS, a gift box was custom printed and hand-delivered. @ad0rnai, who was mere days into her first week at Baseten, sourced a custom metal bookmark, tea from Fortnum & Mason, and created an art deco experience that culminated in an event at The American Bookbinders Museum (planned end-to-end down to the last custom gobo detail for lighting the sidewalks by Ashley). None of this needed to be done. But it was an ambitious idea executed with an extremely high quality bar under a reasonable timeframe and budget. Execution wins across the board. And a no-brainer for leadership to sign off on. People assume the market prizes speed and therefore marketers have to make a tradeoff between speed and quality. It often degenerates to slop to keep up with the Joneses. I believe this is a false dichotomy. I think the market actually prizes value received and human effort. And it's worth taking the big swings. You don't have to make a tradeoff, you just have to be intentional with your choices. But it's important to keep the quality bar extremely high, both when the world is watching your effort and, more importantly, when nobody is watching what you're doing. By the way, none of the above-mentioned people had ever done a project like this. They just kept standards high and figured it out together, thinking how to deliver maximal value to the reader. And the results spoke for themselves. 📷: @katedeyneka

Arjun Naskar ☕️ · @anaskar

How to reduce Claude Code costs by 70% Add `"showClearContextOnPlanAccept": true` to `.claude/settings.json`. When approving Plan Mode, an option to clear all context except the plan will appear at the top, making it easy to naturally avoid long contexts and reduce token usage.

saten@RobotVRM博士 · @saten_work

Insane idea from Opus 4.8

Prompter · @PromptLLM

Proven, Better, New: Mark Pincus on the Rules of Product Innovation

Mark Pincus shares how great founders find winning ideas and why many startups fail. He talks about lessons from building Zynga and handling big challenges. The key is to test fast, focus on what users want, and avoid compromising too much.

good news: /goal in codex has solved all my problems bad news: i'm sitting here twiddling my thumbs while the agents do all my bidding it's no secret i'm a little bit obsessed with codex, and /goal is a big part of that. i put together a 30m ep on /goal including: - what it is - how to write a good goal - how to drive bug zero with @vercel @sentry and codex - two non-technical use cases for goal, including cleaning up 4,000 emails over 4 hours A HUGE thank you to our wonderful sponsor, @mercury - radically different banking loved by over 300K entrepreneurs:

claire vo 🖤 · @clairevo

a little "code mode: react edition" preview! generative ui for any api coming to executor soon

Rhys · @RhysSullivan

Agents answer questions but throw away all your UI. @RhysSullivan of Executor demoed the fix at MCP Night: Agent Mode — 𝗿𝗲𝗻𝗱𝗲𝗿 𝗨𝗜 lets an agent compose your actual product components and deep-link users back into the product.

WorkOS · @WorkOS

the anthropic claude for finance lecture is the best free hour in quant AI right now. bookmark & watch today. It's the most valuable 1 hour in quant AI right now. Then read article below.

Avid · @Av1dlive

Epicure Flavour Explorer

Discover ingredient pairings based on chemical compounds and recipe data

Using AI to write better code more slowly

AI coding tools can help write better code by finding many bugs when used carefully and slowly. Instead of rushing, developers should review and fix critical issues methodically to improve code quality. This slower approach leads to heal…

Launching our new paper on arXiv: we trained the largest multilingual food model ever built. 4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions. All of human cooking compressed into 2 megabytes.

Josef Chen · @josefchen