We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Top Stories
Why it matters: Cheap multimodality and credible evaluation now matter as much as headline capability.
DeepSeek put vision inside the agent stack. DeepSeek says its experimental V4-Flash-Vision-Exp matches V4-Flash on agents, reasoning, and world knowledge, while bringing multimodal-agent benchmark performance close to Opus 4.8. The API supports mixed text/image input across Chat Completions, Messages, and Responses; images cost up to 384 billing tokens each at V4-Flash pricing, and a free Files API supports upload-once, reuse-by-file_id workflows.
Ox Alpha is a market signal, not a settled leaderboard result. OpenCode advertises its stealth model with 1M context, multimodality, zero data retention, near-unlimited use, and claimed capacity for 100T tokens per day; it was available through OpenRouter and OpenCode. A 10-task DeepSWE subset gave it 80%, ahead of Fable at 65% and GPT-5.6 Sol at 52%, but a separate private benchmark found it underperformed substantially. A commentator calls it GLM-5.3 Flash; both the identity and performance claims remain provisional.
Research & Innovation
Why it matters: Progress is shifting toward persistent control loops, richer sensor feedback, and evaluations that measure real task completion.
NVIDIA’s AVO result comes with a benchmark caveat. NVIDIA says its coding agent completed all 183 levels across 25 public ARC-AGI-3 environments without instructions, rules, or stated goals. A monitored account says it learns through trial, observation, and correction while retaining progress across context resets; François Chollet cautions that clearing the public demonstration set is not the same as scoring 100% on the full benchmark.
T-Rex makes touch a first-class control loop. NVIDIA–Berkeley’s method pairs a slow visuomotor planner with a fast tactile expert that corrects motion at four touch ticks per vision tick; its release includes a synchronized 50-hour, roughly 5,500-episode robot-play corpus and tactile-grounded mid-training.
Speech Agent Arena separates sounding good from doing the task. Artificial Analysis compares models with humans across 15 tool-using and 20 non-agentic scenarios. Gemini 3.1 Flash Live Minimal leads preference at 1,046 Elo but has 74.6% task success, while Grok Voice Think Fast 2.0 High leads task success at 94.7%.
Products & Launches
Why it matters: Usable AI is spreading down the hardware stack and into collaborative development workflows.
FreeToken pushes frontier-style local inference onto consumer hardware. UC Berkeley reports GLM-5.2 753B at 14.9 tok/s on one RTX PRO 6000 and Qwen3.6-35B at 39.3 tok/s on an 8GB RTX 4060, with 2–4× Ollama speeds.
Google AI Studio becomes a collaborative repository workflow. Its two-way GitHub sync pushes prompted changes, pulls local or teammate edits, and generates Conventional Commit messages; Google says teams can pull changes and redeploy in under a minute.
Industry Moves
Why it matters: Price, training transparency, and physical capacity are becoming strategic levers.
OpenAI cut GPT-5.6 Sol API and credit pricing by more than 20% for three months, citing capability gains and efficiency improvements. The move makes unit economics an explicit frontier battleground.
Marin opens the training run itself. Percy Liang’s Marin 535B-A23B started on 18.75T tokens, with 80% pretraining and 20% midtraining across 11 GB200 NVL72 systems for about three months, followed by post-training; a four-rung scaling ladder preceded the main run. The project says observers can inspect domain mixtures, sampled documents, live loss, configs, and scaling laws.
Lambda says it deployed 10,368 GB300 GPUs across nine pods and 144 racks.
Policy & Regulation
Why it matters: Provenance requirements are moving from detection experiments into model-provider compliance.
The monitored analysis says the EU Code of Practice requires future models to watermark AI text; Anthropic is rolling out Claude watermarking to everyone, while Google has used the approach since 2024. It says the mark is not human-distinguishable and near-zero-cost, though rewriting can remove it; Anthropic’s FAQ says the detector cannot identify which user generated the text.
Quick Takes
Why it matters: Adoption and embedded workflows continue to broaden beyond standalone chat.
- Codex reached 20M active users; OpenAI credited Codex and ChatGPT Work users with a banked reset while investigating reports of faster limit depletion.
- Runway Ruby converts SDR video to 16-bit HDR in ProRes and EXR for uploaded or generated clips up to 30 seconds.
- Google added Gemini voice controls for Waymo cabin temperature, seating, and route assistance.
AI commentator @teortaxesTex dismissed a benchmark ranking of AI models on agentic coding, saying he has "zero faith" that Ox Alpha is below Qwen 3.8 27B or that Qwen 27B is above GPT 5.6 Sol Max, and suggested the ranking from @bindureddy is a "grift" . He noted that ranking by agentic coding is "more sensible" than other criteria, yielding a chart where "Smaug > Fable" .
A tweet from @MiTypeScript claims a ~60% cost cut for Fable by transparently turning the code into an image and having the model OCR it, with the code at https://github.com/teamchong/pxpipe.
Dan Luu argues that LLMs have removed the cost barrier to performance optimization: the cost of formerly specialized performance work has dropped by many orders of magnitude, so optimizations once reserved for large-scale or lucrative projects are now available to "anyone who can type a few sentences" . His tweet opens with "There's no reason for software to be slow anymore" and links the full essay at danluu.com/perf-opt/ . Marc Brooker is quoted agreeing: "Completely agree with your closing point" . @teortaxesTex amplifies the point: "Enough. We can afford to not have slop anymore" , linking the same essay .
Inkling is being made available for free on OpenRouter, restricted to agentic harnesses, for the next few weeks, with the data collected (disassociated from accounts) used to improve Inkling's agentic performance . The offer was promoted by @andersonbcdefg, who says they actively want to share their data with the company .
Cognition announced GPT-5.6 Sol is now the most affordable frontier model on Devin Desktop and CLI, after OpenAI cut Sol's prices by 20% for the next 3 months; combined with last week's 70% discount, the total discount through October 3 is 76% off list price .
Anthropic continues to build its custom silicon team, per Andrew Curran . Elie Bakouch says "Clive, early employee on the custom hardware oai jalapenos team, joined them a few months ago" and looks forward to "mythos 6 at 1000 tps" . @aidan_clark responds: "Clive is great but lets be realistic about silicon timelines I'm excited for fast Mythos 8" . Bakouch also highlights "CustomChipsKernelBench in the system card" and references "Mythos 5.X release" counts .
@teortaxesTex flags conflicting results for GLM between DeepSWE and Terminal-Bench 3.0, after @TuXinming reported different Terminal-Bench 3 results . He hypothesizes TB3.0 is broader, so GLM's RL training may concentrate in a narrower basin, which would break his earlier hypothesis that GLM had less RL — since RL can expand search space . He still believes Western frontier models have broader RL over more diverse environments and more intense in-domain RL, while ZAI (Zhipu AI) may have relied on intense mid-training specifically around DeepSWE .
AI founders who actually have revenue are now transforming into neoclouds with value-add on top, per @dylan522p .
- DeepSeek released V4 Flash Vision, adding visual understanding to its V4 Flash reasoning/agent model; per DeepSeek's own benchmarks, it comes close to Anthropic's Opus 4.8 on multimodal agent tasks .
- DeepSeek's Wenfeng says he won't pursue "embodied intelligence" before solving hard AGI-relevant problems like continual learning, and expects multimodality to advance before that .
- In an early real-world test, a non-coder used V4 Flash Vision to drive a humanoid robot's camera pipeline hours after release; the model reasoned about spatial relationships, obstacles, uncertainty, and suggested cautious navigation actions .
An AI-focused commentator claims it takes only 2,000–4,000 GPUs to serve 100T tokens/day, assuming a ~300B model, 90% cache hit rate, and an efficient KV cache like DSv4 Flash . They argue modern LLMs can be made extremely efficient and that superhuman-level software engineering doesn't require much intelligence, predicting the public will be shocked once these facts become widely known . Another commentator agreed, adding the impact will extend beyond SWE and hit the public "like a freight train" .
OpenAI is cutting API and credit pricing for GPT-5.6 Sol by over 20% for the next 3 months . @gdb characterized the move as part of a push to offer customers the lowest price on the market for any task while maintaining the highest ceiling on capability .
@z4y5f3, quoted by @teortaxesTex, argues that serving 100T tokens/day would need only 2-4K GPUs assuming a 90% cache hit rate and a ~300B model with an efficient KV cache such as DSv4 Flash ; he adds that modern LLMs can be made extremely efficient and that superhuman SWE work does not require much intelligence, predicting wide surprise once these facts spread .
A tweet reports that Jihao Liu, an algebraic geometer at Peking University, used an LLM to disprove the Yau-Tian-Donaldson conjecture, with the paper on arXiv . The author emphasizes the LLM created innovative mathematical methods rather than combining existing human achievements . The result is framed as a "strong Chinese result in AI for mathematical research," with a caveat about the AI model used .
AI Arena shared first impressions of Qwen3.8-27B with Peter Gostev on YouTube, saying they tested it against models up to 100x its size — DeepSeek v4, Qwen 3.8 Max, Kimi K3, GLM 5.3, Grok 4.6, GPT-5.6, and Fable — on identical one-shot generation and agentic tasks, with scores coming soon .
AI founder friends who actually have revenue are now transforming into neoclouds with value-add on top, per @dylan522p . The sentiment is echoed by @MillionInt: "One day we'll realize we're all just neoclouds with a value add on top" .
Claude Code now lets users start a session directly from their phone: any machine running claude remote-control appears as a device card on the Code tab; tapping it and picking a directory starts a session on that machine . Developer @theo commented that he can't believe Claude is copying T3 Code .
Ox Alpha, offered free by many routers, is actually GLM 5.3 Flash: distilled from GLM 5.3 with more RL (similar to Inkling-Small, Luna), recovering most of the capability; the referenced provider (Z.ai) doesn't supply it directly but bankrolls it, with GPUs overseas . The original observation called it not better than GLM-5.3, more like a post-trained GLM-5.2, and speculated it might be Nemotron 4 or MAI 3 ; @teortaxesTex wished it were Nemotron .
- @bindureddy reports that Ox-Alpha underperforms and is worse than last-generation models: its scores sit alongside Kimi 2.6 (two generations old), and he calls the surrounding hype "pure marketing genius" .
- @teortaxesTex, replying to that eval, shares further results: V4-0813 is "the strongest open model on math, and it's not particularly close"; Flash-0731 is great at data analysis; he adds "I find all this dubious" .
- Ranking by agentic coding, @teortaxesTex notes "Smaug > Fable" .
VC compute strategy analysis (@distributionat, endorsed by @willdepue as "a very good take"): AI-focused VCs should raise 1.5-2x their target fund size and spend the extra on reserved compute facilities allocated to portfolio companies; late-stage funds are already doing this and growth/early-stage funds are starting to adopt it . "If your lead VC cannot find you compute, that's not a lead, that's a co-investor," and LPs should ask VCs their compute strategy . Rationale: AI startups spend most capital on compute, so VCs can gain economies of scale and pricing power; compute will be hard to find next year and likely more expensive, making reserved compute a differentiator as venture capital commoditizes; VCs can bridge the creditworthiness gap between datacenters and startups . Example math: a $200M early-stage fund should reserve ~$60M (30%) for compute, enough at a fictive $4.8/GPU-hour to run a 64-node B300 cluster (512 GPUs) for 3 years ; suggested allocation includes one 32-node training block, one 16-node R&D bridge, three 4-node inference blocks, and four 1-node flex blocks . The author calls this minimal versus neolabs, useful mainly for bridging shortages , and warns finding, pricing, and closing on compute "will get a lot harder in Q1" .
@nptacek argues that frontier model capability is "incredibly jagged" across tasks that are not yet benchmarked, so organizations should assess every frontier model against their own use cases rather than rely on benchmarks or timeline consensus . He points to two models widely criticized for poor eval results that are nevertheless "incredibly" strong in creative writing and spatial work, surpassing the "top" models in those areas .
