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: Open-weight competition is shifting from pretraining scale alone to post-training, deployability, and cost.
GLM-5.3 makes post-training and release governance the headline. Z.ai says its 743B base is unchanged from GLM-5.2 and that gains came from scaling post-training across environments, diverse tasks, and long-horizon workflows. It reports CyberGym at 84.5% versus 77.2%, ExploitBench at 54.4% versus 24.4%, and 105 completed ExploitGym tasks in two hours versus 29. Because the capability is dual-use, Z.ai plans a staged release—controlled partners, broader API access, then complete weights after safety evaluations—and says model-level alignment will accompany the open checkpoint, unlike hosted-only safeguards.
Qwen3.8 makes high-end and single-GPU open weights available together. Alibaba released the Apache 2.0 Qwen3.8-27B, a multimodal dense model with 262K native context extendable to 1M, alongside the 2.4T/95B-active Max model. vLLM reports that the 27B model fits on one Blackwell GPU, includes an integrated speculative-decoding head, and has been verified with tool calls at 1M context. The significance is practical: developers can choose between a very large hosted-style model and a locally deployable checkpoint from the same release family.
DeepSeek V4 Pro 0813 attaches a price warning to the surge. Artificial Analysis scores it 53—eight points above April’s version—and calls it the second-most-intelligent open-weight model it has benchmarked, with roughly 30% fewer output tokens. But new first-party pricing is 264% higher from August 16, lifting cost per task from $0.05 to $0.25 and leaving it only barely on the intelligence/cost frontier.
Research & Innovation
Why it matters: The strongest technical signals are about making long-horizon behavior trainable while preventing agents from carrying failures forward.
Faraday turns research replication into an RL task. Inherent Labs introduced the 27B agent as trained with long-horizon reinforcement learning; the reported system beat Claude Opus 4.8 and GPT-5.5 on held-out paper replication. Its Replica setup uses hypothesis-driven exploration and an automatically generated rubric judge, while the authors say Faraday’s coding-agent tools and rollout analysis point toward scientific capability trained into weights rather than supplied by a complex harness.
Skill libraries can preserve unsafe behavior. A SkillMisevo-Gym study found that all 21 evolved configurations authored unsafe artifacts across 25 agent-method configurations, while 15 caused harm in a fresh session; three malicious tasks raised carryover attack success from 16.0% to 35.3%. Its SafeEvolve wrapper reduced unsafe retrieval by 26.7 points and fresh-session harm by 17.3 points, with only a 0.4-point utility change.
Products & Launches
Why it matters: The agent layer is becoming programmable infrastructure around models, not just a chat interface.
DeepSeek Harness v0.1 is a developer-preview, MIT-licensed runtime in which models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are all replaceable plugins. Perplexity Search SDK brings “Search as Code” to Python applications, letting agents fan out searches and filter, deduplicate, and rank results in code.
Industry Moves
Why it matters: Model competition is pulling product distribution and independent evaluation into the same strategic contest.
SpaceX completed its reported $60 billion acquisition of Cursor. Cursor confirmed the deal’s close and said its team will join SpaceXAI to improve Grok, Grok Build, Grok Bot, the Grok API, and Cursor itself. METR raised around $71 million in commitments for work on autonomous capabilities, recursive self-improvement, monitoring, risk assessments, and incidents; it says it remains independent of frontier AI companies, while acknowledging their significant in-kind token support.
Policy & Regulation
Why it matters: Output provenance is becoming a deployment requirement designed to be invisible to users.
Anthropic says it is implementing text watermarking for Claude to comply with the EU AI Act, alongside other major developers that signed the same Code of Practice. It says the watermark has no practical effect on output quality, adds no hidden characters or extra tokens, costs no more, and cannot be traced to a person, organization, or chat.
Quick Takes
Why it matters: The frontier is also moving through operational benchmarks, product infrastructure, and real-world data quality.
- Gemini 3.7 Flash: On cyb3rops’ THOR benchmark, it scored 72.5%, with 100% threat capture and zero critical misses across 189 real-world findings; the author ranked it first by a wide margin.
- Long-context UX: TokenGremlin reports an upcoming OpenAI upgrade that cut a 741-turn, 231 MB conversation’s load time from 27.6 seconds to 1.66 seconds in an internal test, with 41% less memory growth.
- Document extraction: LlamaIndex’s ExtractBench tested 14 systems on scans, handwriting, and degraded historical documents; the failures did not overlap, underscoring why clean-PDF evaluations miss production blind spots.
DeepSeek Harness, DeepSeek's new open-source agent runtime , is being compared with mature coding agents, but a Zhihu analyst argues that is premature, calling it 'an early infrastructure preview, not a polished Agent product' . In a hands-on test, installation stalled and required cloning 100+ MB of source code, installing 800+ dependencies, and ~27 minutes of setup . Its default toolset is comprehensive (file ops, Bash, web search, subagents, multi-agent workflows, planning) but still 'less refined' than commercial systems like Codex and Claude Code . Fixed context cost is ~7,500 tokens (25 tool schemas ≈6k tokens + system prompt ≈1.4k tokens) . The first coding test hit an execution loop, and the project is a developer preview with compatibility-breaking changes expected . Verdict: promising architecture, ordinary first impressions, not yet a reason to switch from Codex/Claude Code, but it gives DeepSeek a foundation for its own model-native agent ecosystem .
Analyst @scaling01 argues Anthropic's Mythos Preview is likely close to 10T parameters — citing its $125/million output-token pricing (5x Opus, 2.5x Fable), stronger benchmark results than Mythos 5 (UK AISI cyber ranges, GPQA, HLE + tools, OSWorld, most virology tasks, SimpleQA, and an Anthropic missing-reference honesty eval), and compute estimates from the Project Rainier cluster (~5e26-1e27 flops) implying a ~10T model; he also pegs Fable at ~3-5T based on pricing math . @teortaxesTex counters that the evidence is insufficient, noting a 16BA model hitting 80% on ARC-AGI-2 and questioning whether 1-2 evals favoring Mythos-Preview over Mythos justify belief in a 10T model or 'Mythos teacher' . @scaling01 dismisses the theory that Mythos Preview's cheaper serving stems from inference optimization, as it would imply knowingly inefficient training .
DeepSeek released DeepSeek V4 Pro 0813, a new MIT-licensed flagship scoring 53 on Artificial Analysis' Intelligence Index — 8 points above April's V4 Pro, only 1 point above V4 Flash 0731 (tied at 79% on Terminal-Bench 2.1), 7 points behind open-weights leader Kimi K3, with Qwen3.8 2.4T A95B weights released and pending evaluation . It keeps 1.6T total/49B active parameters and 1M context; gains are agentic and token-efficient (128M output tokens vs 185M, ~30% fewer) . First-party API pricing rises +264% blended to $1.32/1M input and $3.96/1M output effective Aug 16, 2026 (old pricing until then; cache hits $0.044/1M with 97% discount; off-peak 50% off) . Cost per Task rises from $0.05 to $0.25, ~20% below GLM-5.2 ($0.32) and barely on the intelligence/cost Pareto frontier . On agentic knowledge work, GDPval-AA v2 Elo hits 1590 (+284, ~84% expected win rate vs April), ahead of Claude Opus 4.7 and GPT-5.5 but below Kimi K3, averaging 39 turns/task vs 23 . AA-Omniscience improves 12 points to +1 (accuracy 49% vs 43%; hallucination rate 95% vs 94%), with ~99% attempt rate vs Kimi K3's 77% attempt and 53% hallucination rate .
X post by @RisingSayak highlights the Hugging Face blog post "State of Open Models" (summer 2026), arguing that short-lived excitement doesn't necessarily lead to impact and that the field should focus on what runs long . Credits to @AdinaYakup, @multimodalart, and @IreneSolaiman .
OpenAI shipped a ChatGPT feature drop (Aug 14): users can auto-generate quizzes with "Quiz me on [topic]", search reservations by natural-language description, paid users can add Google Drive files to ChatGPT Library, and paid users see high-quality suggestions on the home page . OpenAI's @gdb promoted the reservation search feature .
In response to @TenzinTheCyber questioning why a severely overfitted model was approved for release , @teortaxesTex said people misunderstand DeepSeek and that DeepSeek at least doesn't release botched experiments .
Analysts on X are debating Anthropic's model architecture: @scaling01 claims Mythos Preview is a ~10T teacher model, with Model 1 and Model 2 as iterations, while Mythos 5 and Fable 5 are smaller distilled models , later clarifying Fable is likely 3-5T but Anthropic has an internal ~10T model . @teortaxesTex disputes this, saying the served model is ~3T, not 10T , and argues the public evidence is consistent with Mythos-Preview being an undertrained Mythos-5 and Fable 5 a 'defanged finetune' . @scaling01 also cites a pricing disparity (10T at $50 vs 2T at $25) as evidence , though @teortaxesTex rejects the premise .
- An unnamed AI lab is reportedly using AI to accelerate completion of stalled, incomplete drafts of PhD-level research papers, with the resulting work incorporated into its next model release; the claim is hearsay ("I heard about efforts"), with no lab or timeline named .
- @teortaxesTex frames this as a "clever incentive structure": labs buy researchers' working data for pretraining and RL in exchange for ~2 weeks of early access plus free tokens, and says OpenAI already does this "indirectly" .
Ollama fully rolled out DeepSeek-V4-Pro (0813) on its cloud, included in Pro and Max subscriptions; run via ollama run deepseek-v4-pro:cloud, hosted in the US with Zero Data Retention (ZDR) and high performance . Model page: https://ollama.com/library/deepseek-v4-pro.
Teknium is preparing the Hermes Agent GUI Desktop app for a Bots Mode release, now expected Monday, after a series of quality-of-life, optimization, and bug fixes .
TokenGremlin reports OpenAI will ship a performance upgrade for very long ChatGPT and Codex conversations next week . In an internal test on a 741-turn, 231 MB conversation, average load time fell from 27.6s to 1.66s (94% faster), overall memory growth dropped 41%, requests fell from 894 to 16, and loaded transcript items from 15,529 to 64 . The change lets ChatGPT load only the part of the history needed instead of the entire conversation, which should make months-long chats open faster, large-conversation scrolling less stuttery, Codex sessions with hundreds of tool calls lighter, and browser/app memory usage lower .
The community identified the issue behind an AI harness's errors and inaccurate model behavior: the minimal preset hardcoded bash, which Windows doesn't support; the fix is to copy a minimal preset and replace bash with git bash or pwsh . Commentator @teortaxesTex argues the bug led people to wrongly conclude Liang is a "non-CS educated" simpleton over a "non-working sci-fi harness" when the real problem was the Windows environment .
A phishing email impersonating a "HERMES Contributor Update" is circulating; the recipient @noctus91 asked whether it was legitimate, tagging Nous Research and @Teknium . @Teknium confirmed it is a scam and advised not to click anything in such emails, adding that GitHub's @KevinCrosby1 is investigating .
Cartesia's Sonic 3.6 AI voice model adds Urdu and Odia, and improves naturalness across its nine existing Indic languages, including more natural Hindi and Hinglish . The release was announced for India's 80th Independence Day .
- On the All-In Podcast, Gavin Baker said he was told by multiple people that Anthropic CEO Dario Amodei believes Anthropic could become the only private company in the world, with only governments remaining; Baker would "take the under" on that, and David Sacks called the idea "hubristic," adding "This is getting into SBF land" .
- A reply from @_sholtodouglas (speaking as an Anthropic insider) calls the claim "Completely false," saying the source lied to fit a narrative, and states Anthropic is "most worried about economic concentration of power": "There is no world where the government should let any company have that much influence. We need competition and capitalism" .
- The reply adds that the AI market is "literally the most competitive market in the world right now," with every large company focused on cheaper, smarter models, and that success could reduce "the cost of everything to the cost of energy"; it also says "with AGI capitalism gets super weird" .
Alibaba released Qwen3.8-27B, a laptop-sized model now live on LM Studio, pitched as a "frontier-size leap" in capabilities for its size . It requires ~17GB to run locally .
DeepSeek-V4 Pro 0813 beats GPT-5.6 Sol and Fable 5 on DeepSWE software engineering tasks, achieving 88.5% pass@4 at $0.24 per task — 35x cheaper than Sol and 90x cheaper than Fable .
- Per @MaxForAI, someone open-sourced a "dsh-toy" plugin letting DeepSeek Harness (DSH) agents discover, connect to, and control Bluetooth, serial, and USB smart devices — including adult toys. It inserts a unified device-abstraction layer between DSH and hardware so the model never handles vendor-specific protocols; the agent asks for brand/model and auto-routes protocols, falling back to generic Bluetooth discovery when the model is unknown, and hardware complexity is shielded by the plugin layer .
- The project references Chemtrails' device-protocol observations and Buttplug/Buttplug Protocol Specification, and ships a standalone TypeScript implementation .
- The plugin extends DSH's "everything is a plugin" architecture from software pluggables (models, tools, skills, sandboxes, sessions) to Physical I/O — the same structure could later connect robots, IoT, smart home, wearables, and other hardware; the open-source community's first validation was on adult toys .
- @teortaxesTex flagged the project with "It's over" .
DeepSeek V4 Pro API behavior is fueling community speculation that multiple models are routed behind the single deepseek-v4-pro endpoint: users report three distinct chain-of-thought fingerprints (one resembling V4 Pro Preview, one resembling V4 Flash, and a stronger variant), with behavior stable per session . However, TeortaxesTex pushes back: a DeepSeek Harness commit 'fix(preset): align minimal agent with RL composition' suggests the 'god mode 0813' is not a separate model — sandboxed agents used the minimal preset during RL, and 'V4-Grug-Pro needs DSH-Minimal to be awakened' . A follow-up adds that the behavior is 'much weirder and smarter': the team itself may not fully understand how to combine DSH's full affordances with RL .
Alibaba launched Qwen3.8-27B with Day-0 on-device support from MediaTek on the Dimensity Auto Cockpit C-X1 and the latest Dimensity flagship mobile SoC, extending their collaboration from smartphones to vehicles and enabling on-device agentic AI .
