Thariq's triple: the announcement (25553 likes, top of the day): 'We're adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md. You can toggle this behavior in /config'; the reveal (926): 'AGENTS.md support is built off of Claude Code mods, our upcoming way to customize the Claude Code harness. This is a built-in mod, but you'll be able to build custom versions of project instructions yourself'; and the source link (469). Sottiaux's tease (3650): 'We were working on the keynote today with @romainhuet and @sama and most of the fun was trying to figure out how to explain it all to you because there is so much good stuff in there that it's a bit ridiculous all in quick succession. We'll have some things next week already to not keep you waiting.'

Shipper's defense (23): 'there are certainly over the top AI demos on X, but it's very sad to see someone call @jackcheng's work "fake" — He's one of the smartest, most intellectually honest, and craft-focused people I've ever worked with. And his demo is both real and an interesting peak into the future' — plus 'Roose on the loose. Very concerning' (43) and 'THIS CANNOT STAND' (16). Amjad's bare-link post (369), and Nan Yu's pair: 'This is the best movie ever made' (16) and 'The pre-roll was legendary' (0).

Rauchg's trio: the free weekend (1796 likes): 'Free Jev on Vercel. I know what I'm doing this weekend 😁'; the adoption read (711): 'The data and the anecdata on Jev's adoption are shocking. Everyone is adopting it. I think it's a great product, but this is also downstream of the "AI is too expensive/slow" zeitgeist. People are eager to optimize and put AI in even more places!'; and the record day (91): 'Looks like today may be a record day for token volume % of open models on Vercel AI Gateway: 🟦 Open 78.4% 🟨 Closed 21.6%. While spend usually tells a different story, #3 and #4 today are Moonshot AI & DeepSeek. Adding Z.ai, their combined spend surpasses OpenAI (#2).'

Levie's demo (500): 'Jev will be super helpful for agents to make split second decisions in workflows, data classification, judgment calls, and hundreds of other use-cases in the enterprise. Here's a quick demo with Box and Jev… pulls an incident report from Box, asks whether it's customer-facing and how severe it is, moves the file into escalate, monitor, or review folders.' Nikunj's trio: the benchmark (486): 'Jev scoring 3000 kid snacks with multiple criteria in 28 seconds for $0.11 🤯'; the directory (300): 'Built a website called Jevable which showcases all the fun demos on X — filterable by a few categories. Feel free to add your own project!'; and the oops (2).

Peter Yang's Muse trio: the verdict (707 likes): 'Meta's @Muse is the best personal agent I've tried to date. It actually saved me $800+ a year on my cable and phone bills, which is just insane value for a free AI agent… I can see Muse becoming Meta's next billion-user app'; the receipt (230): 'I find this phone transcript from @Muse on how it negotiated $288 in annual savings for my Comcast cable bill super fascinating. Honestly, I think most companies' customer support lines are NOT ready for agents 😅'; and the self-own (35): 'How smart AI gets is inversely correlated with my ability to compose a coherent sentence with no spelling or grammar mistakes.'

Steipete's trio: the team server (83): 'The coolest part: roboclaw runs our team server, is live on Discord, talks with gpt-live and knows about all the sessions it's juggling. We can ask it about context to current and past sessions right while we have meetings'; the deslop tip (50): 'Love having @vhbrzezowski hijacking my sessions and desloping them before the PR lands. Also hot tip: Open the home sidebar and ask your claw to reorganize your sessions'; and the CUA note (13). Zara's law (155): 'It's hard not to create slop when most things you consume are slop. To fix output, first fix input.' Aditya's 'AI is a full circle. Everything old is new again' (6). Plus Claude's weekend prompts (81/81/136).

A brand-new No Priors (published 2026-09-18, new to this digest) with Stefano Ermon — Stanford professor, one of the fathers of diffusion (the 2019 score-based generative models work with Yang Song), now co-founder/CEO of Inception, whose Mercury diffusion LLMs are served in production today. Core threads: the origin (2014 Stanford, generative models over MNIST 'it was hard to publish papers… you had to justify training a generative model as a way to learn features'); the paradigm war — autoregressive (one token at a time, left to right) vs diffusion (coarse-to-fine, iterative denoising; continuous modalities already went to diffusion, discrete text/code is the open question); the killer argument — 'if you think RNNs → transformers was the inflection at training time, the equivalent at inference time is diffusion': autoregressive inference is sequential and memory-bound ('you're spending most of your time moving weights around the memory hierarchy, doing very little arithmetic'), while diffusion processes many tokens at once — a workload that maps to GPUs; test-time compute, reasoning models, and RL rollouts are all inference-dominated, 'so the bitter lesson is that the more parallel solution eventually wins'; the 2024 breakthrough (matching GPT-2-scale autoregressive perplexity with the same parameters, 10x faster) → founding Inception; Mercury today — 'comparable in quality with the speed-optimized models from frontier labs (Haiku/flash/mini/nano class) while being significantly faster'; the Open Call story (voice agents previously on Cerebras custom chips getting the same speed on NVIDIA GPUs via software-level parallelism); economics — 'intelligence per watt, intelligence per dollar'; control as the sleeper advantage (diffusion can be steered from the very first denoising steps against external reward functions — 'easier to control than autoregressive'); data efficiency (denoising = built-in data augmentation — same image, many noisy views); the market math (via OpenRouter taxonomy, 20-30% of workloads are latency-critical today); building the whole stack in-house (custom serving engine — 'you cannot run these on vLLM or SGLang'; own SFT/RLHF/RL stack; keeping IP closed has trade-offs); 50 people + agents ('the bottleneck is more compute than people'); RSI view ('human ingenuity is still super important… we use frontier models a lot, they accelerated our iteration enormously'); and academia's enduring role — flash attention and DPO both came out of his lab: 'there are gems; academia allows you to take contrarian bets.'