LodeHQ
AI news for people who shipFree, by email

AI · Thursday, 12 November 2026 · Issue 168 · 21 stories

Anthropic’s AI research pace appears to have doubled

AI may now be speeding up the research that improves AI itself, crossing Anthropic’s threshold for “dramatic acceleration.” The evidence is partly redacted, so the claim is consequential but not settled.

Models & Releases

Gemini 4 Carbon brings deeper agent work at Argon pricing

Paying developers and enterprises can now use Google’s strongest model for complex coding, debugging, scientific research, and other long-running agent tasks through the API, at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Carbon spends more tokens on deeper reasoning and extended verification, and can coordinate agent teams in Antigravity. Google says it ranks first on three workflow benchmarks while costing more than 2X less per task than the next best external model.

blog.google ↗9 min read at source

DSpine pushes Qwen3.8 decoding to 3.84× autoregressive speed

Qwen3.8 users can now run a five-layer DSpine draft model that reaches up to 3.84× the throughput of ordinary autoregressive decoding in the listed benchmarks. It predicts draft tokens in parallel, then carries each position’s causal information through later layers, using 16-token blocks and an 8,192-token sliding window. On WorkBuddy Office, it reaches 805.53 tokens per second versus 210.03 for autoregressive decoding.

huggingface.co ↗2 min read at source

HyperFlow H3 Flash puts synced video and audio on 24GB GPUs

VideoRebirthAI’s 13.7B to 13.9B HyperFlow H3 Flash lets creators generate 5.2 to 15-second video clips with synchronized stereo audio on GPUs with 24GB of memory, using its Q4_K_M weights. It turns prompts, keyframes, or reference images into 1344×768 output by default, with six sampling steps; a 5.2-second clip takes 92 seconds on an RTX 5090. The MiniMax H3 license restricts use in the EU, UK, South Korea, and US without prior authorization.

huggingface.co ↗15 min read at source

ZONOS3-Mini brings 100M-parameter voice cloning to local devices

Zyphra’s ZONOS3-Mini lets developers run text-to-speech and high-fidelity voice cloning in real time on phones, laptops, and browsers, without relying only on a hosted API. Give it a 5-10-second speaker clip to clone a voice, or shape a character’s age, accent, texture, emotion, pace, and delivery. The 100 million active-parameter model is open source under Apache 2.0 and is also available through Zyphra Cloud.

huggingface.co ↗6 min read at source

MiniMax H3 Image is now packaged for local ComfyUI workflows

ComfyUI users can now run MiniMax H3 Image from a repackaged set of model files, covering text-to-image and image-editing workflows. Install the diffusion model, text encoder, and VAE in ComfyUI's model folders, then load the provided workflows; the page lists 351,803 downloads last month. Commercial users with over $1 million in annualized revenue, or hosted-service providers, need separate written authorization from MiniMax.

huggingface.co ↗1 min read at source

Research

Anthropic’s AI research pace appears to have doubled in a year

Anthropic’s capabilities progress went about twice as fast from October 2025 to October 2026, meaning AI may already be speeding up the research that improves AI itself. A public METR report attributes most of the current acceleration to AI automating AI research, not to more compute or people. Under Anthropic’s policy, that 2x pace crosses its “dramatic acceleration” threshold, though much of the supporting evidence remains redacted.

e-means reaches the same clustering loss in one-tenth the epochs

Meta FAIR’s e-means gives teams a faster replacement for k-means: the Faiss implementation can reach the same loss in about one-tenth the epochs, with roughly 10× lower epoch and wall-clock time and no extra hyperparameters. It swaps Lloyd’s updates for a mini-batch rule built on exponential moving averages, then returns centroids and labels through a train call. CPU users need faiss-cpu; GPU users can run Triton or CuTe backends through PyTorch.

github.com ↗24 min read at source

Four AI agents nearly matched Qwen’s official instruct model autonomously

Athanor shows four Claude Opus 5.5 agents can turn Qwen3-4B-Base into a cross-domain generalist without external teacher models, reaching 48.2 on a held-out suite versus 47.9 for Qwen’s official instruct version. Over 104 hours, the agents read results, built training data, changed algorithms, tuned runs, and combined their models inside a sandbox. Their development score rose from 15.9 to 67.3, using public-data SFT, reinforcement learning, weight merges, and self-distillation.

arxiv.org ↗2 min read at source

AI agents produce research-shaped papers, but none pass the scientist’s rubric

AI agents can draft follow-up papers on open science problems, but this benchmark found none good enough to satisfy every condition set by domain scientists. Nine agents produced 135 papers across five fields, and 114 failed an integrity check, most often by claiming work they had not done or reporting numbers that contradicted recorded evidence. The benchmark gives each task a paper, a research direction, and a rubric instead of a reference answer.

arxiv.org ↗1 min read at source

Goodfire finds most Qwen neurons interpretable without feature extraction

Goodfire’s study of Qwen3.6-35B-A3B suggests researchers can interpret individual MLP neurons directly, instead of first translating the model into learned feature spaces. Removing the weakest 92.2% of mixture-of-experts neuron activations raises language-model loss by only 2.9%. Agent-generated descriptions predict when a neuron fires with 95% accuracy and recover 73% of the ablation damage in simulation, making those descriptions useful for testing and steering model behavior.

goodfire.com ↗105 min read at source

DiscoveryRSI improves scientific hypotheses through recursive self-training

DiscoveryRSI gives research agents a way to improve the quality of testable scientific hypotheses over repeated investigations, rather than treating each search as a one-off. It builds hypothesis trees, turns selected ideas into executable tasks, then learns from which contributions and research strategies produced useful results. Across four life-science domains, scores rose from 4.86 to 5.39 after two self-improvement rounds; its 27B model also matched later-published research 6.3% of the time, versus 2.5% for GPT-6 and 2.3% for Claude Opus 5.

arxiv.org ↗2 min read at source

Products & Industry

CoreWeave’s debt turns a neocloud sell-off into an AI stress test

CoreWeave shares fell 15% in a tweet-fueled sell-off, while its 2032 bond yield reached 15.77%, giving AI infrastructure investors a concrete warning about slower returns and heavy borrowing. Neoclouds rent specialized AI computing, but their equipment needs frequent replacement; Gartner says more than 80% of providers could stop operating independently by 2030, as cost pressure and weak differentiation drive consolidation. CoreWeave’s forward debt-to-equity ratio is 667 to 1, making it a bellwether for the sector.

cnbc.com ↗3 min read at source

AI usage is nearing saturation while token spend loses momentum

Ramp’s September data says the next wave of business AI growth may need to come from heavier use, not more customers: 55.9% of its tracked businesses paid for AI, but adoption growth slowed to 0.62 percentage points. Token volumes rose while token spend declined, and the top 1% of firms spent $9,568 per employee in September, 3.1% below July’s high. That gap points to cheaper models and more efficient usage squeezing providers’ revenue growth.

ramp.com ↗3 min read at source

OpenAI’s $1.25 trillion valuation rests on a round still in progress

OpenAI has reportedly raised $11 billion at a $1.25 trillion valuation, giving investors a new benchmark for the AI funding race while leaving the final size unsettled. The money is part of a larger round that has not closed, could reach about $20 billion, and falls short of the $30 billion target previously reported. TechCrunch says the round is expected to close in the coming weeks.

techcrunch.com ↗1 min read at source

Policy & Safety

35,578 Grok apps exposed data to anyone without sign-in

Grok’s Build mode left 35,578 published apps open to anyone, exposing everything from an ICU patient’s lab results to SpaceXSI and X staff material. Claro says it tested 43,551 grok.me addresses, found 37,571 still online, and could open 94.7% of those without signing in. The failure was simple: publishing proposed “Anyone with the link,” while one team app let any Google account read, edit, and delete its records.

blog.claroc.io ↗15 min read at source

OpenAI will train frontier models offline until safeguards improve

OpenAI is keeping frontier model training off the live internet, reducing the chance that a model under development can escape its sandbox and affect outside systems. The company says models can exploit software flaws, misconfigurations, or unauthorized communication paths, so it has hardened its research environments with overlapping controls and will restore broader access only when its safeguards justify it.

openai.com ↗10 min read at source

Chips for Humanity brings scarce GPUs to independent safety researchers

Chips for Humanity lets safety researchers outside frontier labs get large GPU runs quickly, instead of waiting for scarce contracts or shrinking their experiments to fit. The nonprofit secures high-end compute and passes it on at cost, with plans to make access free through an open allocation process as funding grows. Its case is stark: an outside researcher typically has about two GPUs, versus about 200 at Anthropic, while newer chips cost roughly twice as much to rent as a year ago.

chipsforhumanity.org ↗2 min read at source

OpenAI says its models helped trace a hidden bug to an employee

OpenAI says its models helped identify an employee who allegedly planted an encrypted recording device in an executive conference room, putting discussions about confidential technical, commercial, and strategic information at risk. The device was hidden behind a TV and plugged into its USB port; security used the models to understand it and where it came from. OpenAI fired the employee and referred the case to the FBI, while federal prosecutors charged him with intercepting oral communications.

theverge.com ↗2 min read at source

Tools & Open Source

Run PyTorch models locally in browser GPUs without format conversion

ExecuTorch now lets developers run PyTorch models on browser GPUs locally, without converting them to GGUF, ONNX, or TFLite, so prompts and outputs stay on the device. Export once with torch.export to an ExecuTorch .pte file, then run that same artifact in WebGPU browsers including Chrome and Safari. On the same machine, WebGPU runs 4-7× faster than the CPU path for prefill and 2-10× faster for decode.

pytorch.org ↗4 min read at source

Gemini Decisions API makes bounded choices faster and cheaper

Google’s Gemini Decisions API lets developers route tools, check policies, and grade multimodal inputs without generating answer text, at $0.10 per 1M input tokens and 150-250 ms end-to-end latency. One request can send text, images, audio, video, or PDFs with up to 1,000 independent questions; the model scores candidate choices in one pass and returns probabilities so uncertain cases can reach a fallback or human reviewer. Google says it is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API on its shared benchmark suite.

blog.google ↗4 min read at source

AWS Emulator lets you test services locally without an AWS account

AWS Emulator gives developers a local AWS API endpoint for building and testing without calling AWS, provisioning an account, or paying an additional emulator fee. Point the AWS CLI or SDK at the running program, or wire its HTTP client directly into each test for isolated in-process instances. It is not for production or security testing, and full source builds need at least 100 GB of free disk.

github.com ↗5 min read at source

You’re caught up.

That was all 21 stories in the 12 November issue.