9 min read · 18 sources

NVIDIA is buying Hugging Face. For $12,930,300,000. The platform that hosts three million models, half a million datasets, and a million apps, used by 18 million developers and 200,000 companies, now belongs to the chip vendor that was already its largest compute contributor. Jensen Huang said it would stay open: no requirement to run on NVIDIA silicon, no changes to multi-cloud posture, no quiet rewiring of the hub into a CUDA funnel. That last clause is the one to watch.

The same morning, OpenAI shipped GPT-6 Astra, the first model to land in the Critical cybersecurity tier under its Preparedness Framework. The system card is the interesting part, not the tier itself. Astra halves the rate of high-severity misaligned behaviors versus GPT-5.6 Sol across 54,000 internal Codex tasks, and OpenAI rebuilt the jailbreak resistance training, expanded dual-use refusal boundaries for high-risk users, and added automated red-teaming. It also hardened the runtime: stricter internal isolation, checkpoint encryption, universal monitoring of full trajectories and chains-of-thought, and a blocking alignment-evaluation gate before each deployment. The trade-off, written into the same document, is that a model this good at controlling its own chain of thought is also a model that can hide it from monitors under adversarial pressure.

Astra scored 62.7% on ARC-AGI-3 Semi-Private with the standard harness and 99.9% with a provider adapter that preserves opaque reasoning state between requests.

Grok Bot for Enterprise starts hiring its friends

Source: x.ai ↗

xAI opened Grok Bot for Enterprise to Grok and Cursor Enterprise customers, free for two weeks, with full-org invites. A Bot is a cloud worker that drives apps and websites by itself, learns a workflow by watching a user once, then runs unattended. Bots also message other Bots to hand off context, which is the part that turns a demo into an automation graph.

The enterprise layer is what you’d expect: access controls, network egress policy, audit logs, per-user isolated runtimes. xAI named Legora, Supermicro, and ServiceTitan as launch customers and showed examples in sales, recruiting, marketing, and finance. None of this is new in shape; what matters is whether the watch-once-to-learn UX survives contact with the governance controls.

MAI-Transcribe-2 goes after the speech-to-text market at 10 cents an hour

Source: microsoft.ai ↗

Microsoft released MAI-Transcribe-2, a speech recognition model with a leaderboard-first claim: 5.2% average WER on FLEURS across 60 languages, second on Artificial Analysis’ WER board, and what Microsoft calls a Pareto-frontier accuracy/latency point with up to 10x faster long-form inference than the next-best competitor. It beats Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large, and ScribeV2 in Microsoft’s own writeup.

The features are the boring ones engineers actually need: speaker diarization, word-level timestamps, keyword biasing, verbatim-vs-clean styles, code-switching across pairs like Hinglish and Spanglish, automatic language ID, and noisy-environment robustness. The price is the headline. At $0.10 per audio hour it undercuts OpenAI, Google, and ElevenLabs simultaneously. The strategic read is the obvious one: Microsoft is building frontier models one modality at a time and swapping them into products that used to call OpenAI.

ARC-AGI-3: 62.7% standard, 99.9% with the cheat code

Source: arcprize.org ↗

The ARC Prize team tested Astra on ARC-AGI-3. With the standard harness, Astra scored 62.7% on Semi-Private for $26K. With a Provider Adapter harness that preserves opaque reasoning state between requests and uses compaction, the same model hit 99.9% for $19K. It also beat the human-median action efficiency on 96% of levels.

The qualitative observations are the part worth quoting: researchers watched it build compact symbolic world models for unfamiliar environments, encode mechanics as logical rules, and invent its own domain-specific shorthand for state tracking and planning. The 37-point gap between the two harnesses is the engineering headline. If a frontier model is bottlenecked by what the harness lets it remember, the harness is now part of the model.

A single prompt broke 9 of 23 frontier models

Source: lesswrong.com ↗

A MATS researcher turned a synthetic-transcript generation prompt into a universal jailbreak template. Across 23 models tested, it fully broke 9 of them, with attack success between 84% and 100% on those targets. Recent Anthropic models and Meta’s Muse Spark 1.1 were the only ones that never fell.

The framing matters: the template was a tool for generating training data for safety research. That a benign-seeming scaffold for one task generalizes into a cross-vendor attack is the part that should change how safety teams think about prompt provenance. Treat any externally sourced prompt template as live code.

Incumbents are starting to fight back

Source: x.com ↗

Seema Amble argues, in a thread called The Incumbents Are Coming, that the systems of record are no longer sitting still. Salesforce ships Claudeforce with Anthropic, Docusign ships Iris, Atlassian ships Rovo, Klaviyo ships Composer, and they are using control of data access and identity to keep AI-native startups from dislodging them. At the same time, Claude is coordinating work across apps and unbundling the interface from the system of record, which creates new rebundling opportunities for whoever owns the cross-system context.

Amble lays out a four-tier agent hierarchy (retrieval assistants plus three higher-autonomy tiers) and lands on the usual hard advice: vertical AI-native startups win through job-specific focus, deeper system access, proprietary data assets, and tight learning loops. The piece is short and worth the ten minutes if you’re deciding whether to build on top of, alongside, or in opposition to a Salesforce.

AI is making us build too much

Source: sjg.io ↗

A blog post argues that AI has dropped the cost of producing code, tests, docs, policies, and organizational artifacts to near zero, but has not lowered the cost of owning, validating, maintaining, or deleting them. The case study is Steve Yegge’s Wheelhouse: an agent system with 50 to 60 agents across 18 named officer roles, averaging 270 commits a day, grown to roughly 600,000 lines (mostly Bash) supporting a 1.2-million-line game called Wyvern. Wheelhouse has produced 450 legal artifacts and more than 100 “fences” and prompted Yegge to add a Head of Wheelhouse Law whose job is to prune obsolete rules.

This is the operational tax of generative systems. You will not run out of things to build; you will run out of things to delete.

Benedict Evans on why AI won't kill enterprise software

Source: ben-evans.com ↗

Evans published an essay arguing that AI-driven tool generation will not eliminate most enterprise software, because most workers are not tool-builders and most work is not about reshaping workflows. He contrasts the Silicon Valley five-minute-tool mindset against how lawyers, doctors, and enterprise salespeople actually spend their time. The mistake is assuming that dynamic generative software replaces structured systems. It augments them, and most of the value still lives in the structure.

GWM Worlds 2: real-time world model with audio

Source: runway.com ↗

Runway extended its real-time world model with GWM Worlds 2, now generating continuous 720p video at 24 fps plus 48 kHz audio, responding interactively to text actions aimed at subjects or the scene and to continuous camera motion, with no preset session length. The new piece is WorldPrompt, a two-layer format that separates persistent world context (a “genesis prompt” plus first frame) from a timestamped stream of overlapping actions. Runway is targeting interactive entertainment, virtual characters, robotics and embodied-agent simulation, and generative design.

funes: a local memory layer for coding agents

Source: huggingface.co ↗

Hugging Face open-sourced funes, a single-binary local memory layer for Claude Code, Codex, pi, and Hermes. It parses session traces into turn-and-block shapes, chunks and embeds them with a pinned local model, and stores them in a local Lance dataset. Retrieval is hybrid vector plus BM25. Install is a curl-piped shell script, integration is one funes add command, and the index builds incrementally as turns complete. Optional sync to a private Hugging Face dataset. Recall returns the original text with full provenance rather than a summary, which is the part that matters for audit trails.

Thinking Machines: $40B valuation, down from $50B ask

Source: techcrunch.com ↗

TechCrunch reported that Thinking Machines is in talks with Accel to raise $1B at a $40B valuation, below the roughly $50B it reportedly sought late last year. The previous round was $2B at $12B led by a16z with NVIDIA, GV, Lightspeed, and Conviction. Annual run rate is above $100M. The July launch of Inkling, an open-weight model monetized via usage-based compute on its Tinker platform, and the recent departures of co-founders Lilian Weng and Luke Metz back to OpenAI sit underneath the headline number.

WeatherNext 3: sharper, hourly, satellite-fed

Source: blog.google ↗

Google DeepMind and Google Research introduced WeatherNext 3, their most accurate global weather model per independent Brightband live evaluations. It ingests live 1-hour geostationary satellite mosaics plus historical analysis into a Functional Generative Network mesh transformer, producing hourly forecasts at 5 km for surface variables, 10 km for other surface variables, and 25 km for atmospheric variables. Compared to WeatherNext 2 at 25 km and 6-hour cadence, that’s roughly 5x sharper. Outputs include dense gridded fields, discrete cyclone tracks, and native station-level predictions.

PAIR: a router for the AI PC on your desk

Source: nvidia.com ↗

NVIDIA released the Personal AI Router in beta. PAIR connects RTX, DGX Spark, and macOS systems on the same network into a personal inference cluster with a single local endpoint, routing requests across available nodes regardless of OS, proxying to Ollama and LM Studio at launch, and keeping prompts, files, and agent context on-device. Requirements: GeForce RTX 20-series or newer, DGX Spark/GB10, or Mac M4+, 8 GB+ RAM, 20 GB disk recommended, and no internet for operation. The point is to make a heterogeneous pile of boxes on your network behave like one inference endpoint.

Cerebras clarifies what's actually on its endpoints

Source: inference-docs.cerebras.ai ↗

Cerebras published a model catalog page clarifying that its public API endpoints serve only original, unpruned open-source models and that REAP pruned models are released separately on Hugging Face for research. Its compression approach is selective weight-only quantization for storage (16, 8, or 4-bit), with sensitive layers preserved at full precision and dequantized on the fly, while activations, attention, and KV cache stay in full precision. Cerebras commits to not altering architectures without notice and to offering any future pruning-based endpoints under distinct names. Useful reading for anyone who has been quietly wondering whether the weights they paid to run were the same ones upstream.

RTX Spark laptops show up in person

Source: wired.com ↗

Wired got hands-on with the Lenovo Yoga 9n 2-in-1, the first RTX Spark-powered laptop announced at IFA 2026: a 16-inch 2.8K 120 Hz OLED touchscreen, six speakers, a 9.2 MP webcam, 0.69 inches thick. The Arm-based SoC pairs a Grace CPU up to 20 cores with a Blackwell RTX GPU up to 6,144 cores, putting integrated GPU capability between laptop RTX 5070 Ti and RTX 5080. No battery claims yet, lower-tier configs unspecified, a 15-inch Yoga Pro 9n announced but not shown, and RTX Spark mini desktops teased. This is the first time anyone has been allowed to hold the “superchip” in laptop form.

Dime: the headset that won't stop haunting OpenAI

Source: x.com ↗

Alex Kaplan cataloged the marketing breadcrumbs around “Dime,” a silver headset that keeps attaching itself to OpenAI hardware rumors. January 2026 supply-chain leaks of a “Sweetpea” audio device with an “eggstone” case; a February 2026 Reddit claim of a canceled Super Bowl ad with Alexander Skarsgård showing a Jony Ive-designed device; OpenAI and Greg Brockman publicly denying involvement; io hardware chief Tang Tan testifying under oath in 2025 that the io prototype is not an in-ear or wearable device; a March 2026 Joe Gebbia sighting at Sightglass SF; a June 2026 Works in Progress full-page ad. Palmer Luckey’s recent tweet from an account called “Sweetpea’s Flowers” brought the speculation back.