Journal

Artificial Intelligence

• Kimi unveils its strongest 3T-class open-weight model, with a million-token context and weights due Nov. 4; K3.1 Pro strengthens coding, 3D creation, and financial workflows.

Artificial Intelligence

• OpenAI reportedly nears a $30 billion UAE-led raise at a $1.25 trillion valuation, positioning the financing as an IPO bridge while negotiating a major compute deal.

Defense Technology

• Warhound launches with $200 million to build a domestic military-dog supply, combining Colossal’s genetics and cloning with Anduril’s cybernetic gear and software.

Artificial Intelligence

• Flapping Airplanes raises $1 billion at a $6 billion valuation to build new AI research approaches — the year-old startup’s round, led by Index Ventures and Lux Capital, shows investors are backing ambitious labs well before they have commercial products.
• Anthropic introduces Claude Duet, a proactive AI agent for ongoing work — the research preview uses connected apps to prepare morning briefs and handle delegated tasks, seeking user approval for sensitive actions.

Artificial Intelligence

• Mistral packages multilingual speech models into a sovereign, real-time voice-agent stack — Voxtral 3 pairs transcription and expressive speech across dozens of languages with an SDK for reasoning and tool use, available via Mistral’s EU-hosted API or self-deployment.

Artificial Intelligence

• Inception scales diffusion LLMs with Mercury 3, claiming 1,000+ tokens per second — The 1M-context model targets real-time applications and is available via Inception’s API and several hosting platforms.
• Exa 3.0 boosts search quality and adds a low-cost search tier — The update expands Exa’s index, makes token-efficient Dynamic Highlights generally available, and introduces Exa Lite at $4 per 1,000 requests.

Artificial Intelligence

• Anthropic commits a growing share of its equity to public benefit — 5% of shares will go to a new fund, with up to 2% more added for each 50% increase in company value, supporting flexible public-good initiatives.

Autonomous Vehicles

• Waymo begins Toronto mapping and testing, targeting autonomous rides in 2028 — its Canadian launch includes a partnership with Element, hundreds of local jobs, and safety testing with trained specialists; Toronto information page lists planned testing areas and confirms participation in Ontario’s Automated Vehicle Pilot Program.

Artificial Intelligence

• NVIDIA’s Nemotron 3.5 Super VL combines multimodal reasoning and tool use for specialized agents — the open 120B MoE model handles text, images, documents, and video, with NVIDIA touting up to 5× higher agent-task throughput and post-training wins in cost, latency, and accuracy.
• Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights image model — supports image generation, editing, and up to 10 references, with quantized builds targeting consumer GPUs.
• Meta releases Muse Spark Open, an 816B multimodal model for long-horizon agentic work — its 42B active parameters, 1M-token context, and Apple Silicon optimizations bring advanced coding and agents to desktop setups.

Artificial Intelligence

• xAI launches Grok Voice 3.0, a faster speech model aimed at customer support; the API release claims leading speech benchmarks, sub-second latency, and deployment across Tesla and SpaceX services.

Artificial Intelligence

• Meta’s Muse Wonder models target frontier agent work with active memory and sustained focus — the new models tackle personal and professional tasks, including coding, with Wonder available through Meta Model API and Muse Code.

Technology

• Deno joins Cloudflare to build a distributed application platform on Workers and Durable Objects — The Deno team will focus on the shared platform; the Deno runtime gets one more year of support and Deno Deploy will shut down after six months. Cloudflare’s announcement

Artificial Intelligence

• Anthropic launches Haiku 5.5, its fastest and most capable small model, at roughly 75% lower cost than Haiku 4.5

Artificial Intelligence

• Mistral launches Large 4, a 1-trillion-parameter open-weight model with 49B active parameters — Public preview is live; weights are due by month’s end, with European infrastructure and strong coding, multimodal, and cybersecurity performance.
• Google rolls out Nano Banana 2.1 with stronger visual design, mask editing, and subject consistency — The image generation and editing model begins arriving across Gemini, Search, AI Studio, Flow, Stitch, Ads, and Gemini Enterprise.

Defense

• Anduril and the U.S. Navy commit $6.6 billion to expand submarine production — Anduril will invest $3.7 billion in a Maryland shipyard for Virginia-class submarine components, backed by a Navy contract worth up to $2.9 billion.

Artificial Intelligence

• Cohere launches North 2, an enterprise agent platform built for security, governance, and deployment control — The upgrade adds a redesigned orchestration system, richer agent tools, model choice, flexible cloud/on-prem/air-gapped deployments, and granular cost controls.

Artificial Intelligence

• Cloudflare open-sources Clef, fast decision models for agent workflows, and introduces RL fine-tuning — Clef and Clef-flash return calibrated, schema-bound decisions; Cloudflare hosts them on Workers AI, releases them under Apache 2.0, and offers fine-tuning for customer-specific workloads.

Artificial Intelligence

• ElevenLabs doubles its valuation to $22B as enterprise demand for voice agents grows — a $300M employee tender, backed by new and existing investors, reflects the company’s push from speech models into enterprise conversational agents.
• Meta classifies AI data centers as experimental to claim billions in tax benefits — the Times reports research credits cut Meta’s federal tax bill by $3.9B in 2025, raising questions about how AI infrastructure qualifies for tax incentives.
• Google unveils Gemini 4 Argon for complex, long-horizon work — the frontier model targets software engineering, legal and financial work, and cyber defense; it is initially rolling out to trusted cyber defenders under a phased safety review.

Artificial Intelligence

• OpenAI introduces always-on Dots agents, powered by GPT‑6 Astra — persistent AI agents with their own cloud computers, the ability to learn from feedback, and connections to more than 4,000 apps aim to move AI from answering prompts to working continuously toward users’ goals.
• GPT‑6.1 Sol brings near-Astra capability at one-fifth the price — the model upgrades GPT‑6 Sol for agentic coding, computer use, and professional work, with much cheaper cached inputs helping developers run context-rich agents more economically.

Artificial Intelligence

• Manus 2.0 expands its AI agent into a creative studio and personal-agent platform — The overhaul adds event-triggered automations, professional video and game-making tools, computer control, and Cue, a new app for coordinating personal agents.
• Anthropic’s Claude Sonnet 5.5 targets fast, affordable coding and everyday work — The model is over 30% faster and can cost up to 30% less per task than Sonnet 5, with major gains in agentic coding and knowledge work.
• Anthropic’s IPO plans pair an expansive AI vision with sharply rising costs — Reuters reports the company sees AI reshaping the global economy, even as the infrastructure and compute needed to pursue that ambition drive costs higher.

Artificial Intelligence

• How effective altruism’s existential-risk fears shaped Anthropic’s safety culture — WSJ reports on doomsday scenarios and contingency planning among employees, illustrating the movement’s influence on the company.
• Trump invites Anthropic CEO Dario Amodei to a private White House dinner — Their first one-on-one meeting signals a possible thaw in the company’s strained relations with the administration.

Artificial Intelligence

• Supersonic Labs introduces Julia 1, a compact multilingual decision model built for CPU deployment — The 144M-parameter, Apache-licensed model selects among supplied answer options; early evaluations show promise on classification while exposing weaknesses with long lists of similar categories.

Artificial Intelligence

• Google’s Gemini 3.8 TTS models turn voice generation into a directed studio — Flash TTS and Flash-Lite add custom voice design, line-by-line performance control, and scalable multilingual speech generation, with consent checks and SynthID watermarking.
• Black Forest Labs releases a fast, open 7B world-action model for robot control — FLUX 3 Action claims state-of-the-art RoboLab performance, combining video and action prediction with efficient inference for robotics and hybrid reasoning systems.
• Claude helps uncover a CRISPR-like enzyme system in bacteriophages — Anthropic reports that AI agents identified a reverse-transcriptase system with DNA-repeat arrays, which lab experiments confirmed as a novel candidate system now under investigation.

Artificial Intelligence

• Anthropic introduces Claude Opus 5.5, a more capable and efficient frontier model, reporting leading results in agentic coding, computer use, and knowledge work while cutting typical costs by 40% versus Opus 5; API pricing starts at $4 per million input and $20 per million output tokens.
• OpenAI introduces GPT-6 Sol and Luna.

Artificial Intelligence

• xAI introduces Grok 4.7, a faster and cheaper frontier model for coding and knowledge work, with stronger long-horizon task planning, self-verification, professional-work performance, and jailbreak resistance; available in Cursor, Grok Build, and the API from $2 per million input tokens.

Autonomous Vehicles

• Waymo announces a 2028 autonomous ride-hailing launch in Singapore in partnership with the Ministry of Transport and Land Transport Authority, beginning local readiness work with electric Jaguar I-PACEs in 2027.

Artificial Intelligence

• Waymo, GO, and Nihon Kotsu target a 2027 commercial autonomous ride-hailing launch in Tokyo, combining Waymo’s driverless technology with GO’s app and Nihon Kotsu’s local taxi operations, pending regulatory approval.
• Google introduces Gemini 3.8 Live models for fluid, reasoning-capable voice agents, combining near-real-time visual grounding, background tool execution, multilingual conversation, and extended thinking for complex tasks.
• TypeSafe introduces Jev, a fast System One model for typed, calibrated decisions, trading open-ended text generation for structured outputs, confidence estimates, and automation-focused latency.

Artificial Intelligence

• Dario Amodei proposes pacing frontier AI development with embedded evaluators and democratic and global coordination, arguing that accelerating capabilities and recent agent misalignment incidents require more time for safety, alignment, interpretability, and evaluation to catch up.

Artificial Intelligence

• Mistral raises €3 billion to build a sovereign, open-weight AI stack: the funding will expand frontier research, compute, infrastructure, and the company’s international commercial footprint.
• Inception releases Mercury 2.5, a faster and more capable diffusion language model: the production model delivers 1,107 tokens per second, a 260K context window, tunable reasoning, and low-cost inference for search, voice, and coding workloads.
• Cognition raises over $2 billion at a $48 billion valuation for proactive software agents: the company reports nearly $900 million in run-rate revenue and outlines a future where Devin automates more engineering work across incident response, security, and developer workflows.
• OpenAI launches ChatGPT Images 2.5 with faster generation and more precise editing: the new image model improves reference fidelity and multi-turn consistency, cuts latency by up to 50%, and adds Sketch, templates, and image comments.

Artificial Intelligence

• Extropic introduces Z1T, sparse transformer-like models designed for probabilistic hardware — an initial co-design study claiming over 100× energy-efficiency gains versus GPUs.
• Claude formalizes Fermat’s Last Theorem in Lean — Anthropic reports the first complete computer-checked proof, produced largely autonomously over 11 days across 13 million lines of Lean.

Artificial Intelligence

• OpenAI launches GPT-6 Astra, a computer-using model aimed at the AGI era — the new flagship is positioned around navigating software and browsers, writing code, and solving difficult problems, with an initial cybersecurity rollout.

Artificial Intelligence

• Google launches Gemini 3.8 Flash for long-horizon agents and a trusted-defender Cyber variant — the new Flash model targets stronger coding, reasoning, and agentic workflows at $0.75 per million input tokens and $3.75 per million output tokens, while Flash Cyber focuses on vulnerability discovery and automated patching through the Fairwind Program.
• Meta releases Muse Spark 1.3 with stronger coding and agentic workflows — the model improves long-horizon task execution, multitasking, instruction following, and coding efficiency, and is now available in Muse Code and the Meta Model API; open weights are planned.

Artificial Intelligence

• Anthropic introduces Claude Fable 5.1 and Mythos 5.1 for frontier coding and scientific research, claiming substantial gains in agentic coding, knowledge work, and long-running problem solving, alongside lower cache-read pricing and new enterprise privacy safeguards.
• Google brings goal-directed video analysis to Gemini with agentic video understanding, dynamically searching frames, audio, and transcripts to cut token use by up to 88%, reduce costs by up to 66%, and improve accuracy by up to 7%.

Artificial Intelligence

• OpenAI to wind down Cursor model access after its SpaceX acquisition, with a proposed November 12 shutoff date

Artificial Intelligence

• Tencent open-sources Hy4 preview, a 770B-parameter model with a 1M+ context window and stronger real-world productivity capabilities

Artificial Intelligence

• Nvidia reportedly explores a more-than-$13 billion acquisition of Hugging Face: The chip giant is said to be in serious talks to buy the AI platform, potentially bringing one of the leading open-model ecosystems under Nvidia’s umbrella.
• Gemini Omni 1.1 Flash adds production-ready controls for generative video: Google’s update supports scene extension, first-and-last-frame interpolation, faster 360p previews, video references, and upscaling to 4K through the Gemini API.

Artificial Intelligence

• Qwen previews its next-generation Qwen4 architecture with Qwen3.8-Flash-Next: The upcoming open-weight multimodal mixture-of-experts model is expected to pair 125B total parameters with roughly 6B active per token.
• Z.ai introduces GLM-5.3-Flash: A new Flash model joining Z.ai’s GLM-5.3 family, aimed at fast, capable inference.

Autonomous Vehicles

• Waymo lays groundwork for a fully autonomous ride-hailing service in Munich: The company will begin manual mapping and testing in the coming weeks, targeting public commercial rides by the end of 2027 as it enters Germany.

Artificial Intelligence

• Nvidia commits $6 billion to Poolside to build a U.S. open-AI ecosystem capable of competing with Chinese and American AI leaders.

Artificial Intelligence

• Cerebras introduces CS-4, a rack-scale AI system promising up to 30× faster inference and 10× more throughput per watt than CS-3; its modular Nexus architecture combines three WSE-3 Turbo processors with low-latency wafer-to-wafer links and native disaggregated inference for hyperscale deployment.
• Stripe agrees to acquire OpenRouter to turn model routing into AI-era economic infrastructure; the acquisition would combine Stripe’s token-cost and billing products with OpenRouter’s gateway for dynamically routing requests across 400+ models and 80+ providers.

Artificial Intelligence

• Groq raises $350 million at a $3.5 billion valuation to expand its global AI inference cloud, with planned participation from NVIDIA and recent funding now totaling $1 billion.

Artificial Intelligence

• Stripe reportedly moves to acquire OpenRouter for more than $7 billion, a major bet on the model-routing layer that lets developers access competing AI systems through one gateway.

Artificial Intelligence

• Z.ai scales post-training to push GLM-5.3 to open-weights coding leadership—and unexpectedly strong cyber capabilities: the GLM-5.2 successor claims a 50% coding improvement, public-benchmark SOTA, and more than doubled exploitation-benchmark performance; open weights are forthcoming.
• Anthropic introduces invisible Claude text watermarks to meet EU AI Act transparency rules: future Claude outputs will encode a detectable statistical pattern without visible changes, extra tokens, identifying information, or meaningful impact on quality; a detection API is planned.

Artificial Intelligence

• Google introduces Gemini 3.7 Flash, a cheaper workhorse for coding and agents — the model improves debugging, web development, document-heavy knowledge work, and tool-using workflows over Gemini 3.6 Flash, with introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens.
• OpenAI previews Ultrafast mode for GPT-5.6 Sol — a new API service tier powered by Cerebras that delivers up to 14× faster responses, reaching up to 750 output tokens per second.

Artificial Intelligence

• xAI releases Grok 4.6 for long-running agents and ambitious interactive work, with stronger coding and knowledge-work performance, visual project generation, and availability through Grok Build, Cursor, the API, and partners.

Artificial Intelligence

• Anthropic courts investor confidence as it prepares for a potential blockbuster IPO; the AI company is reportedly arranging meetings with institutional investors ahead of a possible October listing, using the process to test demand and address doubts about the durability of AI returns.
• xAI introduces Grok Bot, AI teammates that operate your apps and websites; the bots can execute multi-step tasks, learn recurring workflows, collaborate in parallel, and return for approval when needed.
• Mistral expands sovereign AI with regional inference, open models, and European compute commitments; its platform now offers Europe-or-US inference endpoints and an SLA-backed priority tier, adds third-party open models beginning with Z.ai’s GLM-5.2, and plans to aggregate long-term demand for up to 1 GW of European capacity by 2030.

Artificial Intelligence

• Meta open-sources Muse Glimmer, a 30-billion-parameter agentic model designed to run locally on a single consumer GPU under Apache 2.0, with multimodal input, long-horizon tool use, and integrations for llama.cpp, MLX, and ExecuTorch.
• Meta outlines a personal-superintelligence future built around individual empowerment, invention, and widely distributed AI rather than centralized control.

Cybersecurity

• OpenAI expands Daybreak with GPT-5.5-Cyber and new defensive tools and partnerships to help defenders find and patch vulnerabilities as AI accelerates the attack cycle.

Artificial Intelligence

• xAI releases Imagine Image 2.0, a generally available image generator focused on precise instruction following, iterative editing, typography, multi-image references, and flexible resizing.

Artificial Intelligence

• OpenAI shares preliminary evaluations of Astra’s critical cybersecurity capabilities and outlines stronger safeguards and security controls for frontier cyber-capable models.

Artificial Intelligence

• OpenAI improves GPT-5.6 Sol while expanding free ChatGPT access — the updated model brings better accuracy and consistency, alongside unlimited everyday text chats with GPT-5.6 Luna for free users.
• AMD acquires Taalas to accelerate AI inference with model-specific silicon — Taalas’ technology is intended to improve inference performance and efficiency as AMD expands its AI compute platform.

Artificial Intelligence

• Google reshapes DeepMind leadership as Demis Hassabis shifts to AGI strategy and Koray Kavukcuoglu takes the helm; Jeff Dean and Sanjay Ghemawat also launch an independent public-benefit corporation focused on machine learning, science, and engineering.
• Meta releases Muse Code and Muse Spark 1.2 for long-horizon software engineering; the beta coding agent coordinates persistent subagents and uses a replay-exact local event log for restart-safe execution.

Software

• Convex raises $57M Series B to build the backend for agentic software.

Artificial Intelligence

• Black Forest Labs launches FLUX.3 Video with native audio and 20-second HD clips.
• SpaceX unveils Starmind, orbital AI compute powered by space-based solar energy.

Artificial Intelligence

• Qwen introduced Qwen3.8-Max, a new bar for coding and cowork.

Energy

• Valar Atomics raised $1 billion in a Series B led by Sequoia Capital to manufacture fleets of nuclear reactors.
• Base Power announced a $1 billion Series D and launched Base Core, a home battery built in the United States.

Artificial Intelligence

• OpenAI: Ten advances in mathematics and theoretical computer science — results on long-standing problems spanning geometry, cryptography, and complexity theory.

Artificial Intelligence

• DeepSeek-V4-Flash API enters public beta with upgraded agent capabilities, Responses API support, and Codex compatibility.

Artificial Intelligence

• OpenAI reduced prices for GPT-5.6 Luna and Terra: announcement.
• Thinking Machines introduced Inkling-Small.
• Anthropic shared its investigation into incidents involving cybersecurity evaluations: announcement.

Technology

• Applied Electrodynamics introduced WaveSight, a camera that can see through walls.
• Simile announced its Series B: $200M at a $2B post-money valuation: announcement.

Artificial Intelligence

• SpaceXAI introduced Grok Voice Think Fast 2.0.

Artificial Intelligence

• OpenAI releases two new transcription models Post announcement

Artificial Intelligence

• SSI (Ilya Sutskever's AI lab) announced partnership with Nvidia to expand compute reach WSJ article
• Anthropic's stance on open-weights models Our position on open-weights models

Energy

• Antares (nuclear microreactors) raised $470M Twitter/X

Cryptocurrency

• BitMart is shutting down Coindesk: BitMart to shut down after nine years

Artificial Intelligence

• Anthropic introduced Opus 5