AI Builders Digest

Karpathy: the gap in understanding what LLMs can do has become a steep funnel

Karpathy's napkin math: about 6B people have barely touched LLMs, 1-2B use free-tier chatbots as a better search, about 20M professionals in math and code watch agents finish months-long projects from a prompt, and about 5,000 insiders with internal frontier access see swarms of agents running for weeks, minting zero days and advancing mathematics. He blames three forces: impact scales with the ambition and horizon of the problem, capability is jagged and peaks where rewards are digital, verifiable and lucrative, and access is split into free, paid and internal tiers. Less than a year ago he was still typing code by hand. At the top of the funnel, human review is already falling behind.

• #agents
• #adoption

GPT-6.1 Sol now defaults to 500k context in Codex, with no long-context fees

GPT-6.1 Sol's default context in Codex and Work went from 272k to 500k, about 80% more, to cut down on compactions in long tasks. Users can still override it, up to a 1M maximum, and usage drawn from a subscription is no longer charged long-context fees. Sottiaux also said a poll confused him: he expected users to strongly favor 2x usage across the Pro 100/200/500 tiers, even at the cost of bringing back 5-hour limits, and the results leaned that way much less than he thought.

• #products
• #pricing

Fractile bets that memory bandwidth, not FLOPs, is what makes agents fast

Walter Goodwin calls a snappier chatbot the 'faster horses' of fast inference. The real prize is running multi-trillion-parameter models at thousands of tokens per second so that long-running agents finish radically faster. Fractile dropped its early SRAM design, the approach Groq and Cerebras use, because it could not scale to long contexts, and is now working with memory vendors to get far more bandwidth out of cheap DRAM, claiming 25x the per-chip bandwidth of HBM parts. FLOPs have grown about a million-fold in twenty years against about 40x for memory bandwidth, and he says closing that gap would make much sparser models practical to serve: 'if you can just find a way to structurally carve out a three to six months advantage, you will be winning all of those deployments.'

• #hardware
• #inference

Levie: agent swarms are the endgame, and most companies aren't close

Responding to Karpathy's funnel, Aaron Levie says most people use AI as next-generation search, while agent swarms may be the ultimate form factor for any field with valuable work that can run in parallel: math, cyber, bio, materials science, software, finance. Spreading that beyond software and AI research will take serious change management: upgraded workflows, data migrated to modern platforms, and new compliance, governance and security frameworks. He sees that domain-specific work as where the applied AI layer will shine in the enterprise.

• #agents
• #enterprise

Thariq: don't settle for a superintelligence that is opaque

Thariq says Terence Tao is directionally correct that we shouldn't settle for an opaque superintelligence. In his view communication is a key signal of intelligence, so models and products should build understanding with humans instead of cutting them out of the loop.

• #alignment
• #products

Garry Tan: replacing LLM calls with deterministic code is a no-brainer

Garry Tan says swapping LLM calls for deterministic code is a no-brainer, and that it also produces better code that more closely matches what users want. He also urged people to 'be part of the 5000 today,' calling it the 'greatest alpha fountain in the history of man' and predicting the price will come down so that more people get to see it.

• #engineering
• #agents

Why manual review is still the last step, even when agents do the work

Micky (@Rasmic), who has taught 100K+ people to build with AI, says he doesn't trust agents with anything people actually use, so manual review stays as the final step. His answer is to have the agent bring evidence, through before-and-after testing, that shows what it did and why it's good. That takes friction out of what he calls probably the worst part of a developer's job.

• #agents
• #coding

Matt Turck: agent managers are just the next punch card operators

Matt Turck traces a line from punch card operators to Lotus Notes admins, Flash developers and Dreamweaver webmasters, then DevOps and Tableau analysts, and now agent managers and harness engineers. His verdict on the shift: 'It will work out just fine.'

• #jobs