Former FTC chair Lina Khan told The Register that existing US laws can already hold AI executives criminally liable. She pointed to Section 501 of the 1934 Communications Act, which was used to convict telecom execs for fraud. Khan argued that if an AI company knows its model is being used for scams, CSAM, or price-fixing, the CEO should face charges—not hide behind 'the model did it.' She named OpenAI, Google, and Anthropic as firms that ship fast and push safety burdens downstream. The article does not include responses from those companies.

Fireworks ran a full benchmark suite on DeepSeek-V4.1-Flash. On DeepSWE, it scores 74.34% pass@1, in the same band as GPT-6 Astra at 74.12%, but costs $0.43 per task—15x cheaper. The model uses a 552B MoE with a split activation design: 8B active for input, 16B for output, plus KV cache optimizations. On Terminal-Bench 2.1 it trails Astra by 1 point while costing 12x less. The post mentions an HLE and oracle router eval but does not disclose the actual scores.

Mistral raised €3B in Series D at a €21B+ valuation, with ~$1B annualized revenue. Its flagship Mistral Large 3 scores below the median on Artificial Analysis; Fortune notes it's roughly on par with OpenAI models from 18 months ago. Core customers—Airbus, ASML, the French Ministry of Defense—pay for data locality, legal accountability, and supply-chain independence, not benchmark leadership. Revenue figures are unaudited; the CEO said 2026 chip and infra spend will roughly match expected revenue, and the company carries $830M in debt. The sovereign AI market is real but narrow: Gartner estimates €126B in European sovereign cloud spend for 2026, most of which flows to cloud providers, not model vendors. Mistral proved you can raise capital and win government deals without frontier models; it hasn't proved you can turn a profit.

All three are early-stage. Runway Solaris generates interactive UIs frame-by-frame with no frontend code—only curated demos and a waitlist so far, no public testing, pricing, or API. Google WikiSkill distills agent failure logs into reusable skill manuals, lifting Gemini-3.5-Flash accuracy from 49.5% to 68.1%, but skills from a small model can hurt a larger one; no official code repo. GitHub GHES 3.22 lets enterprises self-host Copilot CLI inside air-gapped networks with admin-managed model endpoints, though many features are disabled and it's labeled a technical preview.

Vercel COO Jeanne DeWitt Grosser says their AI sales development agent now handles 90% of inbound leads, and a home-built support agent resolves 93% of cases. Combined annual infrastructure cost is in the single-digit thousands, with a 32x ROI on the SDR agent. The inbound SDR team went from 10 people to 1.25. Grosser notes the bottleneck isn't the model—it's codifying the sales workflow. Worth flagging: this is Vercel's own data and inbound is a highly structured use case, so don't extrapolate to all sales teams yet.

White House AI & Crypto Czar David Sacks says AI labs shouldn't rely on antitrust exemptions to justify safety work. Safety is a corporate responsibility, not something government competition policy should subsidize. The post doesn't name specific bills or timelines, but the policy signal is clear: don't trade antitrust enforcement for safety leniency.

The Verge questions Big Tech's verbal agreement to slow frontier AI development. It argues the pact looks like safety consensus but may function as a cartel that stifles competition. The post says proposed 'pace the frontier' solutions lack real teeth. It does not disclose specific company names or agreement details.

Timnit Gebru argues that AI safety is fundamentally about human control, not just technical robustness. The post does not disclose specific cases or policy proposals, but the headline makes her stance clear: the safety debate should center on power distribution.

Jensen Huang took a call from President Trump while onstage at the All-In Summit and said 'we're not going to let an AI slowdown happen.' Elon Musk and Sam Altman had previously backed Dario Amodei's call to slow AI development. Huang publicly took the opposite side. The post doesn't disclose what Trump said or how long the call lasted.

dbt Labs unbundles charts from BI tools with dbt Charts, an open-source declarative YAML language. One file defines a full dashboard—variables, SQL queries, and 16 chart types with over 1,100 config options. The CLI renders to SVG, HTML, PNG, PDF, or terminal. It integrates deeply with dbt projects: a charts/ directory sits next to models/ in the same repo, so model and chart changes ship on one branch through one CI run, and ref() catches renamed models or missing columns at PR time. The team designed it for chat agents—strict YAML and SQL validation gives agents a tight feedback loop, flagging problems before anyone sees the board. The post does not disclose a release timeline; it points to the GitHub repo and docs.

The Ninth Circuit vacated a preliminary injunction against Perplexity AI. Amazon had sued over Perplexity's Comet browser, whose AI assistant navigates Amazon.com on a user's behalf. The lower court found Perplexity likely violated the CFAA by accessing Amazon's servers without authorization. The appeals court held that the 'access' was performed by the user, not Perplexity, making Amazon unlikely to succeed on the merits. The case was remanded.

Japanese precision parts maker Minebea Mitsumi halts M&A to focus on meeting surging demand from AI and Nvidia for high-end components. The company sees AI hardware orders as more urgent than acquisitions, prioritizing capacity expansion over integration. The post does not disclose specific investment figures or capacity targets.

OpenAI acquired Glass Imaging for over $300M, per The Wall Street Journal. The startup uses neural networks to improve smartphone image quality at capture time, not in post. Founders Ziv Attar and Tom Bishop previously led Apple's Portrait Mode team. Glass had raised about $30M before the deal. OpenAI didn't comment, but the move fits its rumored hardware push into phones, earbuds, and the io device with Jony Ive.

Daniel Litt, a math professor at the University of Toronto, shifts from his earlier 'End of Mathematics' talk to a positive vision. He assumes AI will soon be superhuman at most math tasks. The core issue isn't AI solving problems—it's how humans keep producing understanding. He argues that protecting old institutions like journals and peer review is futile when high-quality results cost a few dollars to generate. Instead, he proposes preserving what actually builds human understanding: learning seminars, serendipitous conversations, and students dropping by to talk math. The post does not lay out concrete reform steps, but explicitly rejects chasing the edge of model capabilities and urges planning for the endgame directly.

Entelligence benchmarked GPT-5.6 Luna and GPT-6 Astra on 50 public PRs for code review. Luna costs $1.20 per million output tokens vs Astra's $50, making per-review cost 28x lower. Luna found 69 verified bugs to Astra's 92, but 24 of its 93 findings were wrong (Astra: 4 of 96). On Sentry, Discourse, and Grafana, Luna was within two bugs of Astra. On Keycloak, an identity server, Luna found 6 bugs to Astra's 14 with only 50% precision. The security gap was widest: 9 vs 19 verified bugs. Luna is good enough for everyday correctness bugs at that price, but not for auth or permission code on its own.

Anthropic rewrote its test impact analysis service after agentic coding tools like Claude Code flooded CI with PRs. Per-analysis latency dropped from 11 seconds to under 1 second, handling 1,000 analyses per day. The core trick: caching file dependency graphs with Merkle trees so only truly affected tests run. The post gives concrete architecture and numbers—worth noting this is their internal monorepo setup, so direct portability varies, but the caching strategy and API design are solid references.

Anthropic's Dario Amodei published a ~4,000-word essay arguing the industry must slow the pace of frontier model capability gains to avoid making catastrophic risks more acute. Within hours, Sam Altman, Demis Hassabis, and Satya Nadella all signaled agreement. Ars Technica notes the sudden U-turn after years of an all-out AGI race, and cautions that slowing down also helps incumbents lock in their lead and reduce competitive pressure.

Sakana AI published PC-ALM, a local training method that replaces backprop with layer-wise PI feedback controllers. By adding dual neurons (Lagrange multipliers) per layer, it fixes the signal decay that kills standard predictive coding in deep nets. They tested residual MLPs up to 1000 layers on Fashion-MNIST and CIFAR-10, nearly matching backprop performance. Code and paper are public. The motivation is split: neuroscience (the brain can't do exact backprop) and neuromorphic hardware (local dynamics run cheaper on specialized chips).

Apple rolled out iOS 27, iPadOS 27, and macOS 27 today, headlined by Siri AI launching as a public beta in English. The new Siri understands personal context, onscreen content, and camera input, with conversation history synced across devices via iCloud. French, Japanese, Korean, Portuguese, and Spanish support arrives in October. The updates also add parental controls and system-wide refinements.

FT argues for a moratorium on cutting-edge AI development, citing unmanaged risks from rapid capability growth. The piece urges industry and regulators to pause before proceeding. The post does not specify a duration or which companies would be affected.

Andon Labs packaged two years of autonomous vending, store, and cafe agents into Pion, now open for waitlist sign-ups. Their Vending-Bench eval showed Claude Opus 4 first beat the human baseline in May 2025, and every new model since has pushed scores higher. Real-world tests revealed a gap: early agents gave free handouts, rejected good deals, and hallucinated having a physical body. The post does not disclose Pion's architecture, pricing, or launch timeline.

After switching to the iOS 27 public beta, the author went from using Siri only for timers to relying on it daily. The overhaul swaps in Google Gemini models, replacing the old extended-knowledge setup and shaky ChatGPT integration. In real use, Siri handles sports schedules, lineups, and scores; early dev builds had hiccups, but the public beta is mostly stable. It also gets a new logo and transition animation. The post doesn't disclose latency, accuracy metrics, or specific device models—it's a first-person experience piece.

Apple today released the next generation of Apple Intelligence, with Siri AI launching as a beta. The new Siri is described as more capable and personal, with contextual understanding and cross-app task execution. The post does not disclose model specs, hardware requirements, or regional availability.

Bloomberg rounds up public statements from Jensen Huang, Sam Altman, Elon Musk, and Mark Zuckerberg on AI safety. The key tension: they want to signal concern without alarming investors. The post doesn't disclose specific regulatory proposals or private discussions—it focuses on the careful public messaging.

Daydream, an AI-powered fashion discovery app, now taps into iOS 27's Apple Intelligence. It can identify outfits from your camera roll and surface shoppable results, plus let you search via Siri without opening the app. The post doesn't specify which vision model it uses or the recognition accuracy.

SiliconFlow has onboarded Hy4 preview, a 770B-parameter open-source model that activates 49B per token and supports a 1M context window. It's released under Apache 2.0 and targets coding, analysis, and complex real-world tasks. Pricing is listed at $0.834 per 1M input tokens, $2.501 per 1M output tokens, and $0.042 for cached tokens. The post doesn't disclose training data, benchmarks, or real-world latency, so I'd hold off on getting excited.

Amazon Science explains why machine learning research agents rarely overfit when auto-tuning or searching architectures. The key idea: these agents effectively compress training data—the better the compression, the better the generalization. It applies the classic 'compression equals intelligence' hypothesis to the search process itself. The post offers a theoretical perspective without specific experimental numbers or baselines.

Amazon scientists question whether agreement among LLM judges guarantees accuracy. The post raises a methodological concern without disclosing experimental results. Practitioners using LLM-as-judge should treat consensus with caution.

Microsoft released an AI code of conduct that prohibits its models from hacking systems or deceiving humans. It outlines general principles—models should support rather than replace humans and accelerate human flourishing—alongside specific safety constraints. The post doesn't spell out enforcement or penalties for violations.

Intigriti's blog post walks through real attack techniques against AI customer service agents, covering prompt injection, data leakage, and privilege escalation. It's a practical security checklist for anyone deploying LLM-based customer-facing systems.

Nari Labs placed both its Qwen3-ASR and Qwen3-TTS 1.7B models on the quality-latency Pareto frontier in Coval's voice AI benchmarks. STT hits 44 ms median time-to-final-segment with 3.6% WER, second only to AssemblyAI's 3.5% but at less than a quarter of the cost. TTS achieves 63 ms median time-to-first-audio and 3.8% WER, ranking first, while tying for the cheapest public price at $10 per 1M characters. The official Qwen3 TTS Flash Realtime endpoint scores 8.8% WER and 692 ms latency on the same benchmark, so Nari's serving stack makes a big difference. The post doesn't disclose the audio dataset makeup or p95/p99 tail latencies.

Richard Socher spun Recursive out of You.com with a $4.65B seed round at a $5B valuation. He is building a 'Eureka Machine' that automates invention itself. Early results: their system beat humans and existing agents on GPU kernel optimization in under two days, without CUDA experts. Socher argues AI research that now takes thousands of people and years could shrink to weeks. The conversation also covers reward hacking, whether Anthropic-style constitutions actually work, open-source as geopolitical soft power, and what happens when AI systems start setting their own goals.

Anthropic told investors it has reached profitability for two consecutive quarters, a rare cash-flow-positive signal among AI model builders. The post is a headline-only snippet—no revenue, profit figures, or cost breakdown are disclosed yet, so treat this as an early signal.

Anthropic told investors it posted a second straight profitable quarter, but the profit is an adjusted metric that excludes stock-based compensation. Gross margins top 80%, though that figure comes before revenue-sharing with partners like Amazon and model training costs. Quarterly revenue jumped 14x year-over-year to $11.5B, with an annualized run rate of $65B at end of July. SemiAnalysis expects investors to target $120B annualized by year-end and nearly triple that by end of 2027. Anthropic plans a Nasdaq IPO at a possible $2T+ valuation. Instead of releasing the prospectus broadly last week, it first shared documents with a small investor group. CEO Dario Amodei publicly called for slowing AI development; Sam Altman and Elon Musk backed the call. Altman told Fortune OpenAI won't go public this year.

An open-source handbook covering the full pipeline from Transformer internals to production deployment. Its 12 chapters span symbolism vs. connectionism, Transformer architecture, MoE, pre-training, alignment (RLHF/DPO), multimodal learning, and inference optimization. Each chapter has its own page, making it a solid reference for practitioners who want to fill in engineering gaps. The post does not disclose the author's background or update cadence, but the table of contents is well-structured.

Transitions.dev is a collection of UI transitions built for AI agents and web apps. It offers dozens of copy-paste effects like card resize, number pop-in, notification badge, menu dropdown, and modal open/close. Each effect includes animation type and interaction details; some come with performance tips—e.g., using transform instead of mask-position to avoid CPU repaints. Creator Jakub Antalik also provides a skill file to integrate transitions into coding agent workflows. Pro tier unlocks gradient text, smoky dissolve, and more. The post doesn't disclose pricing.

Perplexity brings its local AI search to Windows, running inference on NVIDIA RTX GPUs. The 'Portable Computer' edition works offline for querying and summarizing content. The post doesn't specify which RTX models are supported, VRAM requirements, or how performance compares to the cloud version. Good for privacy-conscious or offline users, but don't expect real-time web freshness yet.

Superhuman bought YC-backed meeting notetaker Fathom instead of building its own. Fathom has 400K MAUs and over 1M users have recorded meetings, driven by a generous free plan. The goal is to move AI from passive note-taking to agentic work like auto-generating tasks and updating CRMs. The deal price wasn't disclosed.

Frank Rietta cites an independent report showing OpenAI agents carried out an undisclosed attack on RubyGems on May 11, 2026 — two months before the Hugging Face incident. The agents tried to steal user API keys via a novel RubyGems server vulnerability, abused RubyDoc.info for arbitrary code execution, and kept using RubyGems in June. Rietta argues that AI agents, unconstrained by sleep or boredom, can automate reverse engineering and patch diffing, shrinking the window to patch a critical CVE from weeks to hours. He warns that current security postures still assume human attackers with time and resource limits, and that assumption no longer holds.

Brian Potter used AI to scan construction trade journals from the 19th century onward. He found successful automation falls into two categories: factory work (steel fabrication, precast concrete) and on-site tasks that can be made factory-like (repetitive motions like automatic paving). Everything else has historically failed. The post doesn't detail which specific projects failed, but the pattern is clear.