AI Extract launches with SQL-callable PDF extraction

#1 𝕏

Ali Ghodsi says AI Extract was released to extract fields from PDFs, achieving 95% accuracy versus 87% for others at extremely low cost. The function can be called directly from SQL and used throughout the platform.

#2 𝕏

AI at Meta shared new evaluations and demos of Muse Spark 1.2, highlighting multimodal capabilities spanning visual-to-code generation, perception-to-physical action, and audio-visual understanding for video-heavy enterprise workflows. One demo shows the model parsing multimodal observations and calling tools to guide a robot through an unstructured environment to find a rubber duck.

Also covered by: @AI at Meta, @Alexandr Wang

#3 𝕏

Qwen thanked Unsloth for its work and described Qwen3.8-27B as “smaller and sharper than ever,” encouraging the community to try it.

#4 𝕏

bolt.new shared a recap: Two days into building with Bolt, Paul realized he could rebuild EDI software trucking companies have paid $70K a year for since the 1990s; that build is TradeWeave. Eight carriers run on it now, at half the price of the industry standard.

#5 𝕏

Cognition announced Slack Code in Devin, which enables Devin to proactively create dedicated code channels to keep work focused. Cognition also shared details on tuning Devin’s personality in Slack.

#6 𝕏

Madhu Guru shared a recommendation to build a failure-modes taxonomy after v1 of your evals by reviewing the last 500 or 1,000 production interactions and clustering failures under specific names. These categories can inform targeted eval tests and create an improvement flywheel.

#7 𝕏

Aravind Srinivas characterized Perplexity’s Agent API as a full-fledged AI developer platform offering access to frontier and workhorse models, plus tools for deploying them in production workloads.

#8 📝 OpenAI News

Introducing AI Futures - OpenAI published "Introducing AI Futures" on Aug 20, 2026, announcing the AI Futures initiative. The item links to a full article with details about the initiative.

#9 ▶️

DeepSeek is back... and Silicon Valley is terrified

Fireship

#10 𝕏

Santiago called the 2.8T-parameter Kimi K3 the best open-weight model he has tried and described its quality as frontier-level. He runs it on Nebius Token Factory, which provides an OpenAI-compatible API endpoint and a stated speed of 120 tokens/s, though he said he has seen faster responses.

#11 in

Guillermo Rauch announced Vercel Labs’ fx.sh 0.0.5, a tiny, open, native coding agent he says is 10–20x smaller than major coding CLIs, starts instantaneously, embeds in browsers via WebAssembly, and fits in two floppy disks when xz-compressed. He said it would ship the following day with its most-requested, but unspecified, feature.

#12 𝕏

Guillermo Rauch commented that they’re adding versatility to an unspecified platform through backends, queues, workflows, a managed Python runtime, custom Dockerfile support, long-running functions, and Sandbox.

#13 𝕏

Philipp Schmid announced that Awesome Gemma is live on GitHub. It is a list of Gemma resources, tools, and projects, including model cards and collections for 16 Gemma variants, setup guides, fine-tuning recipes, and community tutorials, apps, and demos. People with an additional project, tool, or guide are invited to open a pull request.

#14 𝕏

Boris Cherny recommended running `/usage` for a detailed breakdown, noting that the issue is most often caused by extreme parallelism, a runaway loop, or a very inefficient skill/plugin.

#15 𝕏

claire vo 🖤 recapped this week’s AI workflow switches: “single player claws” to @bot, “mutiplayer slack claws” to @evedev_, “claw mac minis” to remote Codex machines, @DevinAI to more Devin, and “opus slop” to concise mode. They’ll report back on what they regret.

#16 𝕏

Boris Cherny announced that an unnamed item, which had been worked on with customers for a while, was expected to arrive in the fall. He said Mythos-class models require additional safety measures, enterprises need to meet their own privacy and compliance rules, and customers can own and control their data while Anthropic retains none of it.

#17 𝕏

New Fable safeguards for enterprises are being launched to run on enterprises’ infrastructure, providing control over where data lives and who can access it. Developed alongside approximately 100 companies, the safeguards are hoped to roll out more broadly in the fall.

#18 𝕏

Santiago said the key shift wasn’t writing code faster or shipping better code, but building about a dozen tools he otherwise never would have built.

#19 𝕏

Kevin Yien said AI fluency is not best assessed through live coding, highlighting @arjunblj’s proposal: a take-home project using any tools, sharing all traces, and discussing it for 30 minutes. He said this approach uniquely reveals the how and why rather than just the what.

#20 𝕏

Andrej Karpathy commented that a specification could resemble microgpt—scalar-valued Python with for loops—with everything else handled through compilation, characterizing PyTorch as “kind of a crappy IR.”