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: AI’s competitive stack is consolidating around model capability, compute, shared context, and deployment control.
NVIDIA–Poolside ties model production to the chip ecosystem. Techmeme, citing Newcomer, reports a non-exclusive $6 billion licensing deal, a $1 billion NVIDIA investment at a $12 billion pre-money valuation, and NVIDIA job offers to 109 Poolside staffers. The reported package puts model licensing, capital, and talent recruitment in one strategic move.
Muse Spark 1.2 expands Meta’s multimodal agents from perception into action. Meta says the model turns images and video into working code, checks artifacts through rendering and behavior, and can orchestrate a bimanual robot; it also targets audio-visual enterprise workflows. Agent Arena reports net improvement of 2.1% versus 0.9% for Muse Spark 1.1, but the signal is mixed: Bash recovery reached +11.4% and confirmed success +6%, while steerability fell 2.5% and praise versus complaint fell 5.7%.
Slack Code makes agentic coding a shared workspace. Slack’s project-specific code channels expose plans, diffs, and live previews, while high-stakes production pushes require human sign-off. Founding integrations include Claude, Devin, Copilot, ChatGPT, and Vercel agents, with Slack permissions and admin controls inherited by the agents. This moves coding agents from private tabs into auditable team workflows.
Research & Innovation
Why it matters: Capability gains are being matched by attempts to improve internal state, oversight, and evaluation validity.
Activation oracles offer a scalable-oversight path, but not a solved monitor. Transluce trained models that read another model’s internal activations, from 8B to 1.1T parameters, including evaluations of whether coding agents are reward hacking. Performance improved with training and model scale on many evaluations; however, the activation oracle underperformed a full-context monitor on reward hacking, and Transluce says the task remains unsolved.
Recirculation adds inference-time state without retraining. The paper feeds a small part of deeper-layer activations back into a shallower layer, freezing the model weights and adding serial work in prefill but essentially no generation latency. Its abstract reports a 23% perplexity reduction and 21% GSM8K accuracy increase on the Gemma3 family.
Products & Launches
Why it matters: Agent platforms are attacking the operational constraints—round trips, model switching, and decoding latency—that determine production usability.
Claude Platform’s computer use, browser, Skills, and Files APIs are generally available. Claude can now take several computer actions per turn; early-access customers saw 20–40% fewer round trips. The release adds structure-aware browser automation, versioned procedures, reusable files, 500 RPM limits, and 1 TB per organization.
Perplexity’s Agent API offers 41 frontier models from nine providers through one endpoint, with web and finance search, fetch, and sandboxed code execution.
Liquid AI’s DSpark adds speculative decoding to three LFM2.5 models, reporting up to 3.18× throughput on an H100 and nearly 50% lower latency in BFCL multi-tool scenarios, with identical greedy-decoding outputs.
Industry Moves
Why it matters: Application companies are moving beyond wrappers toward specialized models and deployment controls built around their own workflows and risk profiles.
Harvey’s Tenet post-trains a Kimi K3 base with Fireworks on legal, synthetic, and expert data for long-horizon work. Harvey reports relative all-pass-rate gains of 82% on LAB and 22% on LAB Contracts, state-of-the-art on LAB Contracts, and operating cost below one-fourth that of leading foundation models.
Anthropic is preparing a customer-controlled Mythos offering for fall. The company says customers will own and control the infrastructure and data while Anthropic supplies automated safeguards and monitoring and retains no data; an accompanying account says work has involved more than 100 customers and emphasizes monitoring behavior over hours or days to detect coordinated cyber abuse.
Quick Takes
Why it matters: The smaller signals point to where agent reliability, open-model adoption, and financing expectations may move next.
- Self-improvement reality check: A paper finds memory-based agents are highly sensitive to run variance and task order; default orderings can act as a hidden curriculum, and added rubrics only partly recover the degradation.
- Open-model ecosystem: Google says Gemma passed 1 billion downloads and 100,000 variants; deployments highlighted include space operations and India’s 100-million-download Aarogya Setu health app.
- IPO watch: Bloomberg reports Anthropic expects to match or beat SpaceX’s record IPO; a separate monitored report says a filing could come by month-end. This remains a reported plan, not a filed offering.
- Enterprise routing claim: One monitored account says AT&T routes 40% of employee AI usage to open models, cutting coding costs 56% with a 2% quality decline while reserving frontier models for critical work.
The abstract identifies the experimental controls qualitatively but provides no numeric degradation figures for the shuffled-order or repeated-run conditions. The paper re-evaluates two memory-based self-improving methods and broadens evaluation along two axes: multiple runs to quantify variance, and random task shuffling to investigate task-order effects . It reports that evaluation is inherently noisy in complex, multi-step environments and that the self-improving loop can amplify that noise; agent improvement is highly dependent on task order, with default orderings imposing an implicit curriculum . After manually examining the agents' memory, the authors hypothesized underspecification and validated it by adding detailed rubrics and environment feedback to memory construction; this added information only partially closed the performance degradation observed in earlier experiments, and significant gaps remained . The abstract does not report exact remaining degradation values, nor does it detail which variables beyond run count and task order were controlled.
Direct answer: The bundle contains only the arXiv abstract page, so the primary claims are visible, but the method and evaluation setup are not described in enough detail to independently verify them; there is no secondary commentary in this bundle.
- Method claim: Recirculation is an inference-time architectural enhancement for off-the-shelf foundation models that introduces "a specific form of recurrence" to let the model act as a dynamical system for belief-state tracking, motivated by the claim that state updates in feedforward transformers are bounded by model depth. It requires serial processing in prefill but "essentially no additional latency during generation." It is explicitly distinguished from chain-of-thought, popular depth-recurrence/looping techniques, and costly training of recurrent transformers.
- Evaluation claim: The adaptive variant freezes the original model weights and needs only light hyperparameter tuning. On the Gemma3 family, the abstract claims a 23% perplexity reduction on "a suite of datasets," a 21% accuracy increase on GSM8k, and reliable improvements on other downstream tasks.
- Gaps and uncertainty: The exact datasets, task suite, baselines, metric details, and comparison tables are not named in this bundle; the PDF/HTML links are the only route to the full method and evaluation. The 23% and 21% figures are self-reported headline claims from the abstract, not independently verified results.
Hyperscaling by frontier labs is a 'legally fragile strategy,' argues @jd_pressman: no one is interested in granting them Section 230-style protections — tech firms can barely convince Congress to keep existing Section 230 — and hyperscaling concentrates liability into a few firms .
An X post quotes a claim that "We have capacity for 100T tokens per day" and asks who built the model and where the compute came from, linking to an opencode status . @teortaxesTex calculates that at ~80 tokens/s and ~2K tokens/"GPU"/s a 5.3-scale model running at DS-V4 rate would need ~580K "GPU"s to hit 100T tokens/day, so either the figure isn't output tokens or the claim is unlikely . He adds that if the 100T means "all tokens," it is trivial: in serious agentic sessions input+output tokens are under 5% of the total (sometimes 1%), and 1T "honest tokens" a day is just one small cluster .
@scaling01 reports Ox Alpha is a ZAI/GLM model , and speculates it is GLM-5.3 or a larger model — likely larger, arguing ZAI would not stealth-test GLM-5.3 post-launch, and suggesting GLM-5.3's progress comes from distillation from a bigger teacher rather than RL .
Ox Alpha is now available in Hermes Agent through opencode and OpenRouter , and it's free of charge .
Per @teortaxesTex, the mystery model 'Ox Alpha' is likely from the GLM family (Zhipu AI) — speed, time-to-first-token, and cache hit rate match GLM; it is 'clearly not DeepSeek.' Teortaxes expects the release name to be GLM 5.3-Vision or 5.5 with an ~61 score on AA, caveated as strong but unreplicated evidence . @abjevz separately reports 'Ox Alpha' appears to belong to the GLM family . No official confirmation.
OpenRouter released stealth model Ox Alpha, described as a frontier model built for efficient coding, sustained agentic work, and real-world production use, with a 1M-token context window and text, image, and video input . @scaling01 identifies Ox Alpha as a ZAI / GLM model .
AI commentator @saranormous predicts many strong AI companies are about to become desperate because their growth is constrained by compute availability . In reply, @timzaman counters that this compute limitation has been true for a decade .
Steven Strogatz shared an essay making "the case for total opposition to the use of artificial intelligence in mathematics" (arXiv:2608.02859) . François Fleuret critiqued the argument as "bizarre," summarizing it as claiming AI doing math is bad because it only produces results rather than producing more mathematicians .
Ox Alpha, a stealth model by @opencode, is free for the next week with 1M context, multimodal support, zero data retention, generous rate limits, near-unlimited usage, and 100T tokens per day capacity . @theo reports it is "slaughtering all our internal benchmarks" and is surprisingly strong at copywriting, though weak at design .
In a critique of AI-chip startup Etched, @bingxu_ argues the approach is unsound on first principles: on HBM it can't beat GPU/TPU; on SRAM it can't beat Groq/Cerebras; and low-voltage inference is physically impossible because moving data requires energy . A sarcastic reply dismisses such questions, framing Etched's claims as 'wunderkids making physics breakthroughs' and a 'new gpt3 moment' for the insider crowd .
AI researcher @teortaxesTex reports that "Ox Alpha" — an unexpected Chinese frontier model he attributes to Tencent or Xiaomi — outperforms Kimi K3 and is "more well-done than any other Chinese model," particularly in areas where Chinese models are typically weak, and is notably fast . He half-suspects it could be a stolen Claude checkpoint . After more testing he believes it may be the next GLM release with vision, developed in parallel while GLM sat on 5.3, though he cannot confirm its origin . The model's fully visible chain-of-thought mixes DeepSeek-preview, GLM, and Kimi features with fluent English; he argues a full visible CoT makes it unlikely the model stays unreleased, so it is likely to be open-sourced by late 2026 . He stops short of declaring it strictly better than Kimi K3 or GLM 5.3, though it feels that way .
A new paper, "Matryoshka Language Model Suites," proposes training 500M, 1.5B, and 3B models nested inside one architecture and trained jointly . The smaller models are standalone checkpoints, get near-free distillation from the largest model, and share weights plus KV cache for speculative decoding . The paper reports the same performance with 36% less training compute and 14–26% faster speculative decoding . Paper: alphaxiv.org/abs/2608.09703. Practitioner @willdepue calls it "really cool" but is skeptical of the width-mixing approach (thinks fixed width is probably better) and believes the small model should be getting better cross-entropy results than it does .
- NVIDIA paid $6B to acquire poolside's "model factory," which @swyx says has been producing "Thinky-beating models" ("actually not exaggeration, look at the numbers") .
- @eliebakouch: NVIDIA bought only the "model factory" part; many poolside employees (researchers?) received offers from NVIDIA; founders staying at poolside is unusual, and he wonders if poolside becomes a neocloud/compute provider; no mention of PIC (poolside infrastructure company) .
- A reply points to an AI news article titled "poolside gets 12b reverse" for more .
Hugging Face released Diffusers 0.40.0 . Highlights: Modular Diffusers graduates out of experimental ; new model support including LTX2.5, MiniMax H3, Music 3, and Wan Animate 2 ; tensor-parallel support for select models ; new quantization backends SDNQ and Nunchaku-Lite ; and a redesigned test suite that is leaner, better, and sometimes faster . Release notes: https://github.com/huggingface/diffusers/releases/tag/v0.40.0.
A humanoid robot training for the "Robot Olympics" in Beijing ran too fast, failed to stop, slammed into a safety cushion, and broke at the waist . @hardmaru commented on the video: "Winning at all costs" .
Prime Intellect, an AI infrastructure company, announced it is hiring across 25+ roles as it works on "open superintelligence" and infrastructure for self-improving agents . Open roles span research (e.g., AI Research Resident, RL Research Engineer), compute (e.g., Head of Compute, Compute Finance & Strategy), engineering (e.g., GPU Infrastructure, Inference), and growth (e.g., Head of Growth, Marketing) . Applications are open at primeintellect.ai/careers .
AI commentator @teortaxesTex claims Ox Alpha, a Chinese model possibly linked to Tencent or Xiaomi, is frontier-class — better than Kimi K3 and fast, excelling in areas where Chinese models are typically weak; he half-suggests it might be a stolen Claude checkpoint (unverified) . Tech figure @theo reacted with astonishment at the model .
Sakana AI upgraded its translation service Sakana Translate to use the new-generation model Sakana Namazu . Sakana Namazu emphasizes deep understanding of Japanese language, culture, and context, naturally translating uniquely Japanese expressions such as 'お宮参り' into English , and is designed to be user-friendly for Japanese speakers . In internal quality checks using the proprietary TransEvalnia method, Sakana Translate's output was judged superior to all comparison targets in more than 50% of 160 one-to-one Japanese-English comparisons . Try the service at https://translate.sakana.ai; blog post: https://sakana.ai/translate-update/.
