Models & Techniques
Weight-space programming. LLMs from OpenAI, DeepSeek, Anthropic, and others can read and edit neural network weights directly, with replications showing targeted fixes and behavior changes without traditional training.
- GPT-6 Astra Edits Qwen Weights: MATH-500 7.9% to 48.1%(opens in a new tab)
- Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights(opens in a new tab)
- Like GPT-6 Astra, Qwen 3.8 27B can interpret model weights(opens in a new tab)
- I tried MIT's "weight-space programming" on my own use case: Claude rewrote 8 neurons of Gemma 4 12B to change how it translates one Italian tense(opens in a new tab)
- DeepSeek V4.1 Flash 1017 can also do Weight-Space Programming(opens in a new tab)
- What Qwen is thinking when it gets stuck saying "Wait"(opens in a new tab)
- Qwen 3.8 27B on my laptop can interpret model weights, too(opens in a new tab)
- GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights(opens in a new tab)
- [AINews] Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi Shanghai(opens in a new tab)
- Tried the weight editing thing from the MIT post: Some cool applications, but nothing crazy(opens in a new tab)
Hydrogen-1.0. Deep Principle released an open-weight 753B MoE model for chemistry, materials, and life sciences built on GLM-5.3, reporting a FrontierScience-Research win over GPT-6-Astra.
Open frontier models. Tencent's Hy4 and Kimi K3.1 Pro are now available or open-weights, with Hy4 adding computer use and image understanding; Baidu also released a 280B base MoE, and Qwen's code repo hints at future Qwen4 models.
- Alibaba's Qwen Code repo just leaked qwen4-max, qwen4-coder-plus, and qwen4-omni-120b(opens in a new tab)
- Official Statement: Qwen Models Will Stick to Open Source to the End!(opens in a new tab)
- Kimi-K3.1-Pro is now open-weights(opens in a new tab)
- Kimi K3.1 is likely benchmaxxed (with some proof)(opens in a new tab)
- Tencent says Hy4 ranks among the best open models. Its weights aren't out yet(opens in a new tab)
- Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Base Model With 14B Active Parameters(opens in a new tab)
- Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model(opens in a new tab)
- Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models(opens in a new tab)
AI speeds AI research. METR's 11-week review found Anthropic's model improvement rate roughly doubled, with about 80% coming from AI used in its own research, crossing Anthropic's responsible scaling threshold.
Agents & Benchmarks
Proactive coding agent. DeepSeek open-sourced a proactive GitHub coding agent that turns allowed requests into parallel work items, finishing tasks 1.9x faster than a single-session baseline in simulation.
Long-horizon agent benchmark. SemiAnalysis released GDPx, a 206-task benchmark; Claude Opus 5.5 leads at 12%, with GPT-6 Astra at 9.6%, showing agents still struggle with real-world work over hours.
Agent architecture converges. LangChain argues the standard LLM-loop-and-tools agent architecture is converging, with new challenges around long-horizon context, progressive disclosure, and sandboxed tool use.
Codex-scale Git. OpenAI built Pando, a Git backend cutting median push latency nearly 5x for Codex-scale concurrent pushes, and is acquiring Pierre Computer Company, maker of Code Storage and tools that already power Codex.
Meta buys Conductor. Meta acquired AI coding startup Conductor, whose agentic teams let developers spin up projects across parallel workspaces; founder Charlie Holtz says new customers can expect something new early next year.
Infrastructure & Local Inference
Local MoE speeds. Patches like tandem.cpp and Strata plus new quants allow Qwen Flash-Next and DeepSeek V4.1 Flash to run at 97–200+ tok/s on single GPUs with CPU/SSD offload, and 8GB VRAM setups now decode at 50 tok/s.
- tandem.cpp: Qwen3.8-Flash-Next at 97 tok/s on a single 4090 + 64 GB RAM, 3.3× llama.cpp, works for any MoE(opens in a new tab)
- Speedy is shockingly good for qwen3.8 FN(opens in a new tab)
- 8 agents × 250K context, concurrently, on two 3090s (~600+ tok/s total)(opens in a new tab)
- Sorry Sam, my "quite terrible" home DRAM is at 100% now: patched Strata to make DeepSeek V4.1 agent sessions start 20× faster on a good old standard gaming GPU, no datacenter required(opens in a new tab)
- Getting incredible results with a single RTX 3090ti and 128gb of ram. Deepseek-V4.1-Flash-1017 (35 tok/s decode) and Qwen3.8-Flash-Next (130 tok/s decode)(opens in a new tab)
- Qwen 3.8 Flash Next NVFP4 on 64gb RAM and 5090, Basalt. Ditched windows for it. Very hyped for qwen 4 model family now.(opens in a new tab)
- We are in the golden age of Local LLM(opens in a new tab)
- Qwen Flash Next on Strata is unreal.(opens in a new tab)
- Another benchmark has hit the quant war(opens in a new tab)
Quantization competition. Unsloth and ISTA-DASLab are publicly disputing quantization quality for DeepSeek-V4.1-Flash, with Unsloth claiming its GGUFs outperform by 10–20% on long-context coding; the rivalry is yielding better quants.
- Unsloth vs GSQ-RCO - 512K context DeepSeek-V4.1 GGUF benchmarks Terminal Bench(opens in a new tab)
- Kind of enjoying the ISTA (GSQ-RCO) vs Unsloth beef(opens in a new tab)
- Inconsistencies in Unsloth's LiveCodeBench Results(opens in a new tab)
- New Qwen3.8-Flash-Next Unsloth Dynamic V3.1 GGUFs(opens in a new tab)
- Unsloth Qwen3.8-Flash-Next GGUFs V3.1 Update (November 4th 2026)(opens in a new tab)
- Qwen 3.8 Flash Next GSQ RCO quant as good as BF16(opens in a new tab)
GPU resource manager. Berth is a Go-based GPU resource manager giving applications leases, priorities, and a request queue, supporting CUDA checkpointing and parking models in RAM for faster AI workload switches.
Business & Policy
OpenAI raises $11B. OpenAI received the first $11 billion tranche of a round valuing it at $1.25 trillion, below earlier hopes of $30 billion at $1.4 trillion after tech stocks dropped.
Anthropic IPO pressure. Anthropic's revenue grew 12x to $17B but it lost $23.3B due to compute commitments; investors are balking at its $1.5T IPO valuation, with some suggesting $1T.
- Anthropic grows revenue twelvefold but posts $23 billion loss(opens in a new tab)
- Anthropic warns that it’s very close to RSI capability, aka self-building AI(opens in a new tab)
- Anthropic is still grappling with the economics of compute(opens in a new tab)
- Anthropic's $1.5 trillion IPO hits resistance as investors balk at valuation(opens in a new tab)
- [AINews] Anthropic IPO: Whither Now?(opens in a new tab)
- Anthropic pushes ahead with pre-Thanksgiving IPO despite tech selloff, accepts lower valuation(opens in a new tab)
Data center backlash. Voters approved 40 of 41 local data center restrictions, and OpenAI pledged no NDAs with officials and to pay full electricity costs, while Meta warned AI expansion could slow.
- After the midterms, a bigger clash over data centers could be coming(opens in a new tab)
- Big Tech claims voters’ AI backlash is “anti-Trump” not “anti-tech”(opens in a new tab)
- Texas was the new epicenter of the AI data center buildout. That might be kaput.(opens in a new tab)
- Data centers lost the midterms(opens in a new tab)
- How a Blue Wave midterm might reshape the AI industry(opens in a new tab)
- After the midterms, is the AI industry in for a rough two years?(opens in a new tab)
- Big Tech lost big in the 2026 election(opens in a new tab)
- All the latest updates on AI data centers(opens in a new tab)
- After the midterms, things are looking rocky for Elon Musk and data centers(opens in a new tab)
- OpenAI might rebrand its data centers as ‘AI compute campuses’(opens in a new tab)
- OpenAI is stepping back from NDAs for data center projects(opens in a new tab)
- After midterm rout, OpenAI tries to win over communities it needs for AI expansion(opens in a new tab)
- Meta warns that election backlash against data centers could make AI expansion harder(opens in a new tab)
- ‘Total repudiation’: The blue wave is about to hit AI(opens in a new tab)
- OpenAI is already trying to shape the 2028 election(opens in a new tab)
That's the week in review – see you next Sunday.