We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Karpathy: the capability gap is now a funnel
Andrej Karpathy revisited his April post and argued that the gap in understanding of LLM capability is widening into "a sharp funnel." His rough numbers ("napkin math") break down as follows:
- About 6B people have barely touched LLMs.
- 1–2B are casual free-tier users.
- About 20M professionals in math and code see agents finish multi-week projects from a prompt.
- About 5,000 people have internal access to frontier systems .
Karpathy describes what that last group sees as swarms of thousands of agents working for weeks on software, cyber, science and math. He says "human review and comprehension are starting to fall behind," and points to the 722 frontier-math manuscripts as an example . He gives three causes: how ambitious the task is and how long it runs; capability peaking in domains that are digital, verifiable and economically valuable, because of RL on verifiable rewards; and access tier . This framework helps explain why reactions to AI progress this week varied so widely.
Mathematics absorbs the OpenAI release
Le Monde, via a widely shared post, reports that Fields medallist Hugo Duminil-Copin was "deeply shaken" by OpenAI's release of more than 700 AI-generated manuscripts, which include work on 16 questions directly relevant to his research. He is quoted saying, "They have annihilated my field of research" . Fellow Fields medallist Jacob Tsimerman thinks mathematics will soon rest on proofs no one understands. He says that is already true of many of his own theorems, which build on results whose proofs he doesn't know .
Hugging Face turned OpenAI's problems into open-source RL environments. Clément Delangue calls them "early and rough": the verifier accepts only the exact original formalization, so an equivalent proof can score zero . A slide from Terence Tao's "Math 2.0" lecture at Caltech set off a separate argument about whether a cure needs to be understood if it works . François Fleuret, for example, noted that many drugs have unknown mechanisms of action .
Agent-assisted open research is also moving fast. An integer-multiplication effort crossed κ = 10⁻³, with much of the day's progress coming from optimizing and canceling existing parts of the algorithm . Of 151 PRs opened in 24 hours, 89 disclosed using Codex and 48 Claude. Codex produced the most advances, but the top teams used both models .
Frontier model updates
- GPT-6.1 Sol context: the default context in Codex and Work went from 272k to 500k tokens, about an 80% increase. The goal is fewer compactions. Users can set it as high as 1M, and OpenAI says it will go higher .
- Long-horizon behavior: one researcher reports that GPT-6 Astra followed requested work durations precisely, but only for tasks over 2 hours. Asked to work 72 hours, Astra stopped at exactly 72h while Fable 5.1 gave up after 3h. The researcher asks whether stopping to check in might actually be better .
- Anthropic rumors (unconfirmed): users say requests appear to be going to a new Opus 5.5 or 5.6 checkpoint that "looks really, really good" . Others say Fable 5 is routing to a possible Fable 5.5. One commentator speculates it will be released ahead of Anthropic's IPO . Anthropic has not confirmed any of this.
Open weights: cheaper inference, specialist fine-tunes, and a margin debate
Two new inference providers launched on DeepSeek V4.1 Flash:
- River AI prices DeepSeek v4.1 Flash and GLM-5.3 Flash 20% below DeepSeek's and ZAI's own offerings. It credits a training and inference stack it wrote from scratch in Rust .
- Dolphin claims "Opus 5-class" coding at 97% lower cost than frontier models .
Brian Armstrong says Coinbase fine-tuned Qwen3.5-9B on internal data for fraud prevention, and that it beat frontier models at lower cost. Coinbase has added a fine-tuning skill to its internal harness, and Armstrong expects some backend services to be replaced this way . The company's write-up projects fraud-prevention savings of roughly 3× the inference cost .
Jaya Gupta argues that inference is a "permanent reverse auction." In her view it is weakly defensible because switching providers is "one line of code" and neolabs keep adding spare capacity . Sarah Guo replies that inference is "a scale game" and that leading providers rarely lose customers . In a SemiAnalysis clip, the speaker argues that even the labs underestimated their own demand and signed GPU contracts at high premiums. They call it "a revenue and profitability fast takeoff" rather than an AI one .
Small "decision models" kept arriving after last period's launches from Microsoft and OpenAI:
- vllm-sr released Decision 3.0, open weights from 0.6B to 27B, and claims #1 on both Jev Decision Index boards in its own evaluation .
- Liquid AI released open-weight d1-3B and d1-omni-600M, which make decisions from images. Liquid reports 35 ms per image on Jetson AGX Thor and is targeting on-device checks for robots .
Multi-agent research
Meta Superintelligence Labs' GitSwarm gives hundreds of asynchronous agents one shared Git repository, with no central planner. With GPT-5.5 Codex workers it solved all 30 IMOProofBench-Advanced problems in one run. It scored 79.4% on ProgramBench, against 65.1% for a single agent at a similar budget, and the score rose as workers were added. 70.1% of failed research commits were cited by later accepted work .
Sakana released new detail on MASS, its self-improvement loop where one model acts as solver, optimizer and evaluator. Training only on task-solving data also improved the optimizer and evaluator roles. Multi-agent traces made more efficient training data than single-agent ones, and optimizer capacity can limit how fast it improves . Sakana warns that when a model is its own evaluator, shared blind spots can let mistakes be "accepted as progress" .
Other signals
- Meta Muse retention: above ChatGPT at day 7 (>30%) and just below it at day 14 (22%) . Scott Stevenson notes the curve keeps falling instead of flattening, and guesses users switch off its proactive notifications .
- XPeng's Iron humanoid: a post says mass production is beginning, with a European rollout in H2 2027 .
- Baidu's Unlimited OCR: a 3B open-source model that parses 100-page PDFs in one pass . Jerry Liu's independent check puts it at about 46.2% on ParseBench (#115), noting it skips visual elements and the most complex tables .
- Packaging: Cu/diamond composites debuted at Semicon Taiwan at about 800 W/m·K, roughly double pure copper's ~400 W/m·K. The main obstacle is bonding copper to diamond .
- Andrej Karpathy argues that people’s experience of LLM capabilities is increasingly divided. His rough estimates range from ~6B people with little exposure and 1–2B casual free-tier users to ~20M professionals using frontier models in math and coding; he estimates ~5,000 people have internal access to frontier systems. He describes the latter as seeing agent swarms take on major software, cybersecurity, scientific, and mathematical work, while human review and comprehension struggle to keep pace.
- @hrishioa distinguishes three agent-app approaches: online-only agents trade execution access and capability for portability and ease; locally run systems trade security and ease for raw capability; and integrated products such as Claw, Dots, and grokbot aim to avoid that trade-off but, in his view, poor configurability and one agent handling too many roles increase cognitive load and cap complexity.
@dejavucoder says software engineering has changed significantly but is not over, while claiming that coding is now automated; the post names no specific system or scope, so this is a broad assessment rather than a concrete capability announcement.
SoTALab says it built Blend3DForge, a Blender-agent training dataset that captures the full creation workflow rather than only final renders: briefs and revisions, planning and code, Blender operations, failures and fixes, editable project files, and QA evidence. One example records 305 tool calls and 729 trajectory events; its 68-item rubric evaluates visual quality, dimensional correctness, clean rebuilds, and editability, with 58 critical criteria required to score 5/5 and a minimum 90% visual score.
Dan Shipper’s AI-era team-structure view highlights “two-slice teams” . One team reports naturally settling on 8 engineers across 4 major independent product/research streams while taking on high-value work, without deliberately choosing that ratio .
A study of neural PDE solvers on two-dimensional Kolmogorov turbulence found that lower trajectory-prediction error can coexist with less accurate learned dynamics: some models had larger local errors but amplified existing deviations less, and one Fourier Neural Operator’s local residuals exposed problems before aggregate error distinguished a trajectory that later became unstable. Models could match the overall spread or event frequency while getting spatial patterns or event-triggering conditions wrong; in one persistent extreme-vortex case, 10 of 11 configurations missed the reference event group, while the remaining configuration frequently predicted events in the non-event group. The team recommends evaluating error accumulation, responses to perturbations, and conditions for physical events—not prediction accuracy alone—and cautions that the findings concern this turbulence system.
Meta Muse was reported to retain over 30% of users at day 7—above ChatGPT—and 22% at day 14, slightly below ChatGPT . Scott Stevenson cautioned that Muse’s retention curve keeps declining instead of flattening, and hypothesized that users may disable proactive notifications around weeks 3–4 .
T3 Code nightly added a proper /goal flow and visual callout.
A user reports using an LLM to find bookkeeping transactions missing receipts, access the corresponding vendors, download receipts, and attach them; browser use worked without an API and was described as flawless and fast enough .
@oneill_c speculates that AI regulation could center on FDA-like scrutiny of reinforcement-learning environments, checking for inconsistencies, mistakes, and reward-hacking pathways that could teach models misaligned behavior .
Meta chief AI officer Alexandr Wang says he brought an approach he learned from Elon Musk—small, exceptionally talented teams with little hierarchy and room to execute—to Meta, where Mark Zuckerberg embraced it as central to the company’s AI strategy.
The poster says they may be able to use GPT-6.1 Sol Ultrafast as their main model on the $500 plan while just barely hitting their limits . They clarify that the 583/week figure is total usage, not Sol-only, and that their Opus run used a $200 plan; a second run found more prompts, but most apparent omissions were noise from broken PR monitors, and ts-rust was excluded as a long-running task rather than day-to-day use .
A proposed AI screen-time-control approach would replace gameable rules and filters with cheap decision models: an always-on agent would interpret goals and take over a computer as needed. This is an idea, not a product announcement; a responding user said they wanted it on their phone.
A user reported that Hermes lacks the “ultra fast” catalog entry for GPT sol 6.1 and rejects messages over 1 MB; they also described errors as intermittent. Teknium said they would implement compression but doubted it was the underlying cause, noting that image uploads may still exceed 1 MB and compression would only help some affected requests; they suspected a possible service-side bug.
@TheZachMueller said they have four SSD-hybrid inference setups running, with RAID10 mirroring across cards in pairs; the quoted post described SSD-hybrid inference as arriving, but neither post gives model or performance details.
Karpathy says the gap in shared understanding of LLM capability is widening into a sharp funnel: his rough estimates are about 6 billion people barely exposed to LLMs, 1–2 billion casual users of free-tier chatbots, 20 million professionals using frontier models in math and code, and 5,000 people with internal frontier access. He describes the frontier tier’s preview as thousands of agents collaborating over weeks on major software projects, cyber operations, science, and mathematics, with human review and comprehension beginning to lag. He attributes the divide to task ambition and time horizon, uneven capability in digital, verifiable, economically valuable domains, and access tiers.
@badlogicgames describes AI coding as “pressing the magic button” and cargo-culting, with results ranging from abysmal to good and little understanding of how to improve them . The account argues that junior developers should still learn fundamentals, while agents can guide them when they are reluctant to ask senior colleagues .
@NandoDF argues that AI’s superiority in mathematical proofs could help advance understanding of the universe and address healthcare and energy problems, while also bringing anxiety about meaning, finances, social impact, and uncertainty. He expects machines soon to write AI papers more effectively and notes the difficulty mathematicians and students may face as AI takes over work they have devoted themselves to. A linked interview raises a related proof-comprehension concern: Fields Medalist Jacob Tsimerman thinks mathematics may soon rely on proofs no one understands, and says he already builds on accepted results whose proofs he does not know.
Coinbase fine-tuned the open-weight Qwen3.5-9B on proprietary data for fraud prevention and reported that it outperformed frontier models at lower cost; an online experiment projected fraud-prevention savings at roughly three times estimated inference cost.
Jaya Gup argues inference companies can still raise large rounds by pointing to data-center and compute needs, while neolabs renting out capacity and companies pivoting to pure inference add supply and intensify recurring price undercutting. Because app companies can switch providers with “one line of code” and move substantial spend to cheaper options, Gup sees inference as weakly defensible despite its high valuations. A reply disputes the low-stickiness thesis: it calls inference a scale game, says leaders rarely lose customers (and that churn is to other leaders), and emphasizes workloads, financing, and operational and technical capability. Another reply warns against selling $20M of inference at negative gross margin without a path to scaling supply.
One user reports that Claude Next created an unspecified output from scratch using Blender MCP during a two-hour run prompted by one sentence, with no ready-made model; the user called it their most impressive model output so far. This is an anecdotal capability signal, not a reproducible evaluation.