Karpathy dusts off an April tweet (30115 likes, top of the day): the gap in shared understanding of LLM capability is *widening* — now a sharp funnel rather than two groups talking past each other. ~6B people (~75%) have barely touched LLMs; ~1-2B (~20%) are casual free-tier users treating them as a better Google; ~20M (0.2%) see first-hand that weeks/months of professional math&code projects now complete with a prompt ('less than 1 year ago, I was writing code manually by hand… occasionally pressing Tab'); and ~5,000 (~0.00006%) with internal access see swarms of thousands of agents collaborating for weeks on software mega projects — minting zero days, machine-speed cyber, new science and math — while human review falls behind (people still doing 'archeology' of the OpenAI-HF incident; mathematicians poring over 722 frontier manuscripts). The funnel is driven by: 1) impact scales with ambition/problem size/horizon; 2) jaggedness — capability peaks in digital, verifiable, economically valuable domains; 3) access tiers. Levie's echo (343): the vast majority use AI as a next-gen-search chatbot, 'dramatically underrepresenting' full capability, while agent swarms are possibly the ultimate form factor — but a huge amount of change management (workflows, data, compliance) is needed, which is exactly where the applied layer will shine. Garry Tan (1113): 'Be part of the 5000 today. It's amazing. The lucky thing is the price will come down and more people will see this… Greatest alpha fountain in the history of man.'

Sottiaux's Day 7 triple: the context bump (10608 likes): 'We have increased default context length for GPT-6.1 Sol to 500k when used through Codex & Work, up from 272k or an ~80% increase. We will go higher in the future, but it helps reduce occurrence of compactions and with tasks that require keeping a lot of context and nuance over periods of time. It was carefully tested and is the new sweet spot for most of our users. You can still override it… with maximum context length of GPT-6.1 Sol of 1M'; the pricing note (2553): 'With this change, we are not charging long context fees for GPT-6.1 Sol when drawing usage from the subscription'; and the poll postmortem (4625): 'The results of this poll has gotten me confused. I definitely thought it would sway much more favorably in the 2X direction across Pro 100/200/500 at the cost of bringing back 5h limits in order to smoothen the demand curve.'

Thariq on transparency (1219): 'I think Terrence Tao is directionally correct, we shouldn't settle for a super intelligence that is opaque. A key signal of intelligence is communication. We should strive to make products and models that build understanding with humans instead of disintermediating them.' Rauchg's pointer (2659): 'So cool. The future of learning.' Garry Tan's engineering rule (1624): 'Replacing LLM calls with deterministic code is a no brainer and also results in better code that better matches what users want.' And Turck's job-evolution meme (54): 'People were punch card operators then they were Lotus Notes admins, then they were Flash developers and Dreamweaver webmasters, then they were DevOps and Tableau analysts, and now they're becoming agent managers and harness engineers. It will work out just fine.'

Peter Yang's pair: the father-daughter build (75): 'Building a @modretro game with my daughter and Codex'; and the trust quote from Micky (@Rasmic), who's taught 100K+ people to build with AI (24): 'I'm not at a point where I trust the agent, especially with stuff that people are using… The manual review is like the final step. All the work's done, and I have as much evidence from the agent as possible. With evidence-driven testing, before and after, and even greploop, we're trying to build trust. The agent presents evidence: This is the work that's done, and this is why it's good. For anyone who's been a developer for any time, manual review is probably the worst part of your job. This sort of helps minimize the friction.' Zara Zhang's usage note (93): 'The only place where I use Gemini these days is in Chrome.' Shipper's side project (73): 'made a feed that automatically tracks the books im reading (and have read going back to 2012!) maybe we should have a feed for everybody @every could be fun to see.' Swyx's prep (36): 'reupping my talk tips thread for all our lovely AIE NYC speakers. see you tmr at speaker dinner!' — plus the minor notes (Peter Yang's YouTube-comments quip at 3, Garry Tan's SF politics take at 239, Nan Yu's 'Who's gonna tell him' at 0, Nikunj's 🔵👼 at 20).

A brand-new No Priors episode (new to this digest) with Walter Goodwin, founder and CEO of full-stack AI chip company Fractile (founded summer 2022, ~150 people). Core threads: the 'extraordinary zoo' of AI silicon is mostly identikit — hyperscaler internal efforts (Google TPU, Meta MTIA, Microsoft Maya, OpenAI Jalapeno) all realized through a small number of ASIC houses like Broadcom, sharing HBM, Tensor Cores and TSMC packaging, leaving a dearth of teams going all the way up and down the stack; the two core technical bets — inference from day one (when a recently-IPO'd 'inference' chip was a training chip until 18 months ago), and memory bandwidth: after two years on an SRAM-based design (like Groq/Cerebras) abandoned over scalability (context length and KV caches grow), Fractile spent two more years on skunkworks with memory vendors to get extreme bandwidth to high-capacity DRAM — a platform ramping H2 next year that unites GPU-like capacity with Groq-like speed; why speed matters — the snappier chatbot is 'the faster horses of fast inference' (Henry Ford), the real elevator is radically faster long-running agents, and data-center economics collapse to cost per gigabyte of memory; bandwidth scaling laws — 25x more bandwidth per chip than HBM parts unlocks sparser MoEs (1/16 → 1/128/1/256) and FLOP-cheaper attention, a multiplying factor on global throughput (FLOPs scaled ~1M-fold in 20 years, bandwidth only ~40x); AI-forward chip design — compressing the observation-to-volume latency is the chip-space equivalent of frontier models ('carve out a 3-6 month advantage and you win all deployments'), and on intent-to-GDS2 in 10 years: divide by four and subtract a bit, with NP-hard place-and-route loops and surrogate-model approximations as the bottleneck, 'we're gonna be generating a lot of reasoning tokens before we place a given circuit'; and market structure — first-party chips as NVIDIA price leverage, frontier labs' premium-speed raison d'être ('otherwise we'd all be using Kimi models'), and why third-party frontier chip players sustain for decades: an all-in proprietary-silicon bet could kill a lab in the nine months before a rival's chip-only breakthrough lands.