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 competition is moving beyond model scores toward the physical capacity, local availability, and operating economics that determine deployment.
NVIDIA and OpenAI are locking in the physical substrate of scale. At Ohio’s PORTS-Pike campus, NVIDIA and SB Energy are securing land, power, and shell for NVIDIA compute, with OpenAI as tenant; OpenAI will operate a full-stack DSX AI factory. NVIDIA says initial capacity is 4.25 GW, with a possible 3.75 GW extension, while OpenAI’s existing and planned commitments represent about 12 GW through 2030—expandable to roughly 16 GW and $600 billion of NVIDIA compute. The support covers defined lease and power payments plus residual value, not the full site cost, and phases in as facilities come online from 2028 to 2030. This moves NVIDIA upstream from selling chips toward securing the sites and demand that support repeated upgrades.
Local models are moving from impressive demos to distribution. Cline reports that Qwen3.8-27B matched DeepSeek V4-Pro and GPT-5.6 Luna on the Artificial Analysis Intelligence Index—the first local model it says has reached frontier capability. Unsloth says its GGUF build reached 2.7 million Hugging Face downloads and became the platform’s #2 trending model. The combination of a frontier-level third-party score and rapid downloads makes local availability a competitive signal, not just a hardware hobbyist story.
Agent evaluation is becoming task- and cost-specific. After analyzing more than 1.7 million Agent Mode sessions, Agent Arena added cost-per-task/Pareto views and Code, Chat, and Work categories. The leaders diverge: GPT-5.6 Sol leads Code, while Claude Opus 5 variants lead Work and Chat.
Research & Innovation
Why it matters: The strongest technical gains this period come from changing how models are scheduled, trained, and instructed—not simply enlarging them.
Weave treats idle time as a cluster resource. The scheduler time-multiplexes multiple rollout/training jobs while preserving each job’s on-policy synchronization. In a production-scale evaluation with 328 H20 rollout GPUs, 328 H800 training GPUs, and 200 heterogeneous RL jobs, it reported 1.82–1.99× higher throughput, 1.84× lower provisioning cost than naïve disaggregation, and 100% SLO attainment. Its trace simulations reached decisions for 2,000 jobs in 591 ms versus more than five hours for brute force; the approach is aimed at PPO, GRPO, DAPO, and similar phase-dependent workflows, not fully asynchronous systems.
SocialRL shows the value of training for delegation, not generic pleasantness. The paper’s 4B model was trained across six social environments; on held-out negotiation scenarios it matched or exceeded the GPT-5 family per domain, with 78% of buyer openings anchoring below target versus 3% untrained. Its unified model reached 0.627 average utility across the six environments, above GPT-5.1’s 0.619 and GPT-5.2’s 0.613. These are task-specific results from a v1 preprint, not a claim of general 4B parity.
Agent skills appear to stabilize procedures more than supply knowledge. A paper finds procedural anchoring accounts for 65.7% of cases where skills help, versus 4.5% for explicit knowledge injection; actual-use precision falls from 29.6% to 3.3% as the pool grows from five to 100 skills. In a separate 87-task SkillsBench test, Gemini 3.7 Flash rose from 44.9 to 65.9 and #8 to #2 with curated skill files, at roughly $1.80 per task and 186 seconds—beating the reported cost and latency of Opus 5.
Products & Launches
Why it matters: Agent products are increasingly being embedded inside the software and interfaces people already use.
Gemini 3.7 Flash’s Android computer-use quickstart uses an ADB screenshot → model planning → normalized-coordinate action loop, without accessibility trees, element IDs, or XPath. The open-source implementation works across native apps, webviews, and dynamic canvas interfaces, positioning the model for UI testing, bug reproduction, and task automation.
Cursor’s Origin is live. The code-hosting platform supports repository hosting, pull requests, review, and deployment alongside Cursor’s agents; GitHub repositories sync bidirectionally and remain the source of truth. Vercel, Buildkite, and Depot integrations are already available.
Industry Moves
Why it matters: Commercial AI is showing both accelerating revenue concentration and unusually large new capital commitments.
Anthropic’s scale is accelerating ahead of its IPO. Bloomberg reports that its annualized revenue reached $65 billion by the end of July, more than seven times its end-2025 run rate; its latest completed quarter reportedly exceeded $11.5 billion versus $787 million a year earlier, with positive adjusted operating income.
Capital is still clustering around infrastructure and interfaces. The Rundown lists Higgsfield’s $400 million Series B at a $5.4 billion valuation, Groq’s $350 million Series A at $3.5 billion, and Wispr’s $280 million Series B at $2 billion.
Quick Takes
Why it matters: Small product changes show how agent control, pricing, and multi-agent coordination are becoming operational features.
- Perplexity Computer now gives each connector an Allow, Always Ask, or Deny setting; recurring runs inherit thread approvals.
- OpenRouter cut GPT-5.6 Sol pricing by 50%, to as low as $1.25 input/$7.50 output per million tokens on flex.
- Hermes Desktop’s Bot Mode gives each agent its own role, model, memory, skills, tools, and inter-bot communication.
- Seedance 2.5 leads Video Edit Arena and now offers native 1080p output with 10-bit color.
Abstract-level verification of arXiv:2608.14036. The paper claims:
What skills do: Skills work when noisy trajectories become procedural anchors that stabilize execution; procedural anchoring accounts for 65.7% of skill cases vs 4.5% for explicit knowledge injection, so skills stabilize action rather than inject missing facts.
Performance: Skills improve over Workflow Memory by 6.06 points in matched comparisons.
Skill-pool size effect: Retrieval is a separate bottleneck—as pools grow from 5 to 100, actual-use precision falls from 29.6% to 3.3%.
Distractor robustness: Confusable distractors impair offline identification, yet downstream success remains stable; exact ground-truth invocation is neither sufficient nor necessary.
Failure conditions: Skills fail under brittle assumptions, incompatible contexts, or insufficient adaptation.
The abstract also reports a consolidated taxonomy of three high-level categories and twelve skill-use modes from 8,135 controlled trial records (238 valid unique labels from 240 open-coded records).
Gap: This bundle contains only the abstract; the full paper's experimental detail is not available, so these claims cannot be independently verified beyond the abstract.
The abstract reports that SocialRL is a general training recipe that trains social reasoning directly, applied to a 4B model across six domains; the reported 4B-versus-GPT-5 negotiation results are per-domain matches/exceeds on negotiation games and a unified-model average-utility comparison. Verification is limited because only the abstract/metadata are in scope.
- Method: SocialRL is described as "a general recipe that trains social reasoning directly," applied to a 4B model, with every domain trained in-domain under the same recipe. Two consolidation strategies are named: cascade RL and multi-teacher on-policy distillation (OPD).
- Evaluation domains: Six domains are listed: Deal-or-No-Deal, CaSiNo, Craigslist, Job Interview, Calendar, and Marketplace. Every policy is evaluated on all six.
- Reported 4B-versus-GPT-5 negotiation results (in-domain specialists): On held-out scenarios, the 4B "matches or exceeds the GPT-5 family per domain," closing 73-122% of the baseline-to-frontier gap on the negotiation games; 78% of buyer openings anchor below target versus 3% untrained. The phrase "per domain" and "on the negotiation games" limits the claim to those games, not all six domains.
- Reported 4B-versus-GPT results (unified model): A unified 4B reaches 0.627 average utility across all six environments, matching or exceeding GPT-4.1 (0.625), GPT-5.1 (0.619), and GPT-5.2 (0.613).
- Caveats and scope limits: Results are self-reported in the abstract; no error bars, significance tests, or benchmark details are provided in this bundle. The paper is a preprint (v1, 13 Aug 2026) and the submission history confirms only one version, with no peer-review status stated. The abstract additionally notes that cross-domain transfer is game-structure-dependent and that a theory-of-mind scaffold helps only through training, which qualifies the generality of the method.
A user (@antoniogm) claims that uploading and logging everything into OpenAI Finance makes it a better CPA and financial advisor than any human he has paid, and that it similarly outperforms human lawyers and doctors; he argues humans remain involved only due to offline-world inertia . In a follow-up, @ericmitchellai urges trying it, saying it delivers hundreds of dollars of value for $20 a month .
OpenAI is hiring for AI safety/misalignment work, with Micah Carroll calling it "the best time ever to join" and describing the team as "bandwidth bottlenecked" with support for impactful misalignment work . Hiring topics span monitoring, misalignment and monitorability assessments, misalignment science, supporting third-party auditing (citing a recent Redwood collaboration), and communicating risk externally via system cards, blogs, and safety cases . An RSI/misalignment researcher role is open in San Francisco .
Zhihu launched the online roundtable 'Embodied Intelligence: The Next Evolution' (具身智能，走进现实), running Aug 17–Sep 8, 2026 as WRC 2026 and the World Humanoid Robot Games approach . The series brings together academics and industry leaders from Tsinghua, SUSTech, Wuhan University, and robotics companies including Realsee, LinkerBot, Genisom AI, D-Robotics, Leju Robotics, and ECOVACS . Planned discussions cover how robots are evolving, remaining barriers between lab demos and industrial deployment, and where embodied intelligence goes next .
US datacenter electricity use is estimated at 217 TWh vs China's 115 TWh annually, a lead smaller than the AI compute disparity would predict . This contrasts with China generating more total electricity than the US .
Together AI added native A/B testing and staged rollouts to its endpoints, letting customers split live endpoint traffic into one control and up to 20 variants with fixed percentages, ramp via a single call, and remove the experiment to send 100% of traffic back to control with nothing to unwind . The company says customers are using this to safely and rapidly roll out new weights across small to large deployments . It positions endpoint-level A/B testing as superior to shadow traffic and client-side flags: clients keep the same endpoint name, API, and keys, with no feature flags or hash-mod-100 code . Full walkthrough: https://www.together.ai/blog/a-b-test-models-in-production.
Together AI announced endpoint-level A/B testing for production models: a live endpoint can split traffic into one control and up to 20 variants, each with a fixed percentage; experiments ramp with a single call and deleting them returns 100% of traffic to the control . The post argues shadow traffic only proves operational soundness, not user preference, and that A/B testing belongs at the endpoint (same endpoint name, API, and keys) rather than in app code with feature flags . Full walkthrough: https://www.together.ai/blog/a-b-test-models-in-production.
Someone claims to have unlocked the true power of Deepseek-V4-Pro-0813 with a simple harness that fixes thinking-process errors, completely outperforming Fable across every task with "insane" benchmark scores; a GitHub link is shared . Caveat from @teortaxesTex: the harness is not an actual J-Space-utilizing or J-Space-informed tool (which would be worthwhile), but a pure harness-side workaround; if that is enough for such gains, it is arguably even more damning for Deepseek .
Fashion house founder @yanabana is using OpenAI's Codex as her 'technical co-founder,' putting AI to work across the entire atelier so she can focus on creative work; she shared the approach in a video titled 'This is How I AI' . @npew highlights the use of Codex for fashion design .
Qwen3.8-27B, a local 27B model, matched DeepSeek V4-Pro and GPT 5.6 Luna on the Artificial Analysis Intelligence Index — the first time a local model has scored frontier capability . Alibaba Qwen responded that this is 'just the beginning' and the model will keep expanding into more fields .
An X post by @teortaxesTex notes that the DSH repo is the #1 DeepSeek repo "and it's not particularly close," while the next two repos are "world-historically significant" and DSH is a "barely usable harness prototype" .
@peterom predicts "Fable at Home" — a ~30B parameter model with capability similar to larger models — will be available sometime between January and May next year . @scaling01 cautions that the plot supporting this trend is misleading: it reflects "benchmaxxing" and a lack of hard benchmarks to distinguish the best models, not genuine capability progress .
@TheTuringPost's must-read papers of the week: AI4AI at Test-Time (strong-to-weak capability transfer) , Persistent Recursive Worlds (autonomous software evolution) , AgentRewind (recoverable execution for long-horizon agents) , Demystifying Agent Skills , Envs-FORGE (environment synthesis for agent RL) , Stealing Reasoning Traces from proprietary LLM APIs , Agent Safety as a runtime contract , WorldClaw (agentic 3D world generation) , Second Thought (parallel reasoning for agents) , ScienceFlow (long-horizon agent for ML research) , and OmniScientist (omni-modal AI scientist) . Full list with links at turingpost.com .
AI researchers released LittleLearner, a 5B parameter model trained from scratch on LittleCurriculum, a corpus restricted to K–5 material, to test how LLM behavior changes when training data never exceeds fifth-grade content; an interactive chat demo is available .
The 'Pacing the Frontier' AI governance effort reportedly hinges on getting China on board: the 'Pace faction' — including many leading figures, but notably not Zuck or Elon — has one month to secure President Trump's support and craft a proposal Xi might actually consider . Commentator @teortaxesTex pushes back, arguing Xi is not the leading Chinese AGI/tech expert, the fate of open AGI won't be decided in a day, and September will only confirm an accepted mode of dialogue for the next stage .
Gemini 3.7 Flash scored 92% on a physical tool-use benchmark (saturating it), up from 32% for Gemini 3.6 Flash released three weeks earlier — described as a step change in LLM robotics capabilities.
Sakana AI's Namazu model is now available on Vercel AI Gateway, with support through the Vercel AI SDK for integration into web applications . The model powers the new Sakana Chat and offers reasoning tailored to Japanese and business contexts .
According to @gdb, the most overlooked skill in machine learning is creating evals; 'worthy metrics which beg for improvement are the root of progress' . Echoing this, @yacinelearning observes that many who went on to work on frontier AI started by contributing to a challenging eval .
Google DeepMind's Gemini 3.7 Flash ranks #2 on SkillsBench with a score of 65.9, just 0.1 points behind Grok 4.5 (66.0), a gap inside error bars . The benchmark, from BenchFlow, runs each model twice on the same 87 software tasks in one OpenHands-CLI harness — once with default tools, once with curated skill files (reusable procedures from xlsx/pdf/pptx utilities to seismic model selection, power-flow, and PCAP workflows) — and measures the skills' value from the gap .
Without skills, 3.7 Flash scores 44.9 (#8, behind GPT-5.5 at 49.9 and Grok 4.5 at 48.8); with skills it gains +21.0 points vs a +14.9 average, the largest gain of any top-10 model (Grok 4.5 +17.2, Opus 5 +15.2, GPT-5.5 +12.3), moving it to #2 . It uses skill files slightly more than average (2.4x vs 1.9x per task), while GPT-5.5-Codex opens them most often (4.8x) but scores 3.3 points lower, suggesting instruction application matters more than reading .
Cost/latency: 3.7 Flash runs at $1.80/task and 186s, outscores Opus 5 by 5.5 points, which costs $2.51/task at 505s; skill files add effectively no overhead ($1.80 vs $1.85, 186s vs 190s) . It is flash-tier priced at $1.50/$7.50 per M tokens . Full results: https://www.vals.ai/models/google_gemini-3.7-flash