Daily AI Briefing
The 79 stories that mattered in AI, curated and summarized from dozens of sources by AIBriefs.
NVIDIA will be the exclusive AI compute provider at SB Energy's PORTS-Pike campus in Ohio, with OpenAI as the tenant for 8 IT-GW. NVIDIA invests $1.5B in SB Energy and provides credit support for the initial 4.25 IT-GW, with an option for an additional 3.75 IT-GW.
Grok Bot is in early beta for Grok Heavy, Cursor Ultra and Cursor Team Premium users, with a 7-day free trial on desktop. Each bot gets its own cloud computer and signs into Gmail, CRM and websites without an API. SpaceXAI's launch is the latest bid by Elon Musk's company to keep pace with Anthropic and OpenAI.
An unreleased research version of Claude increased the proven lower bound for the fraction of Riemann zeta zeros satisfying the hypothesis from 41.6% to 67.2%. It tested 650 ideas across 60 sub-agents, with results validated by Anthropic mathematicians and formalized in Lean.
Databricks closed $5 billion in financing at a $190 billion valuation. CEO Ali Ghodsi said the capital will go toward Genie, Lakebase, and Unity AI Gateway, with the company benefiting from the agentic AI wave.
SoftBank is investing $200M in Gravis Robotics' Series A — claimed the largest in construction robotics history. The ETH Zurich spinout retrofits excavators with its Gravis Rack autonomous control kit and uses synthetic training to bridge the sim-to-real gap.
Dean, Google's chief scientist for 27 years, co-founds the company with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. Backed by Radical Ventures and Khosla Ventures — with Kleiner Perkins, Lightspeed, and Doerr Capital — it plans to automate complete experimental loops to accelerate science and engineering; Alphabet will take a stake.
Anthropic's annualized revenue run rate passed $65B at the end of July, up from $47B in May and $9B at end of 2025; Q2 revenue topped $11.5B, up 14-fold year over year. The company is expected to IPO this fall seeking a $2T+ valuation, which would be the largest market debut on record. Rival OpenAI's run rate is about $40B.
FCC added humanoid and quadruped robots to its Covered List, blocking new foreign-made models from US equipment authorization on national security grounds, a move targeting China's ~85% share of the humanoid market. Power inverters are also covered; existing devices keep working.
OpenAI bought back $7 billion in employee shares, valuing the lab at $852 billion — the same as its March round, which raised $122 billion. Bloomberg reported the deal; OpenAI filed confidentially with the SEC in June for a potential IPO, but the tender suggests that debut may wait.
The tender offer has been in the works since OpenAI closed its record $122 billion funding round in March, and the company is weighing a public listing.
Bloomberg Businessweek reports Anthropic is negotiating to buy startup Decart for $6 billion. VC Rudina Seseri argues leading AI firms like OpenAI should focus on better data over better models.
Bitcoin miner Riot Platforms has struck a $9 billion, 20-year compute deal with Anthropic. The agreement is the latest sign of miners pivoting to AI infrastructure.
OpenAI released GPT-5.6 Cyber, a model trained for security work, available through Daybreak Red, a new tier in its gated cybersecurity program. Internal tests showed it answers 95% of advanced threat queries. The model is also available on AWS via Amazon Bedrock.
NVIDIA released NemotronLabs VoiceChat 11B, an open 11B end-to-end speech-to-speech model with ~450 ms turn-taking and live tool calling. It performs streaming speech understanding and generation in one unified network, supporting transcription, translation, sound recognition, audio Q&A, TTS, and full speech-to-speech.
MiniMax-H3 ranks #1 among open models in Video Arena for text-to-video and image-to-video, tied for #3 overall. Open weights dropped with Day 0 ComfyUI, fal, and Vercel AI Gateway support, offering multimodal input, native audio, instruction-based editing, and up to 2K output.
Moonshot AI's Kimi K3 reportedly outperforms top US models at a fraction of the cost, and its free weights are pushing OpenAI, Google, and Anthropic to reconsider closed releases. Open-weight models let developers inspect, self-host, customize, and build without vendor lock-in, though they aren't fully open-source.
AgentCore Payments middleware lets LangChain agents pay for paid APIs: on a 402 response it checks the price against a session budget, signs the payment, and retries. Built with AWS on Amazon Bedrock AgentCore, it uses the Coinbase-created x402 stablecoin micropayment protocol, with every transaction traced in LangSmith.
OpenAI is removing text-chat limits for all ChatGPT users, powered by the new GPT-5.6 Luna model, which becomes the default for Free and Go users. Plus and Pro users get an upgraded GPT-5.6 Sol model and a thinking slider; internal eval shows 62% fewer factual errors for Luna and 68% fewer for Sol vs GPT-5.5-Instant.
OpenAI flagged ChatGPT messages from 25-year-old former Goldman Sachs analyst Darren Zhou, who detailed plans to rape and murder his ex-girlfriend, and reported him to the FBI. Court records quote Zhou: "I'm gonna kill her by the end of this month."
Amodei argues AI concentrates power on its own due to capital-intensive training, rejecting the choice between regulation and decentralization. He endorses pre-launch vetting and says real accomplishments will earn public trust.
Model ML uses GPT-5.6 Sol to automate financial research and analysis, generating editable PowerPoint decks and Excel workbooks.
An AI model solved the Jacobian conjecture, a problem open since 1939, verified by mathematicians within a day. Anthropic employee Levant Alpöge announced the result, drawing over 20 million views on X.
Jeff Dean is leaving Google after nearly 27 years, along with Sanjay Ghemawat, Oriol Vinyals, and two other top AI scientists, to found Discovery Loop. Google will take a stake in the startup, which aims to automate the scientific method across fields like biology and chip design.
DeepMind released Gemini 3.5 Flash Cyber, a specialized AI model built on 3.5 Flash that discovers, validates, and patches software vulnerabilities. The model will be exclusively available to Google's security clients.
Anthropic CEO Dario Amodei said it will be possible to cure most human disease in ~5-10 years, including cancer. Critics like Gary Marcus and Eric Topol push back, noting no AI-designed drug has reached clinical adoption yet.
On May 20, 2026, OpenAI announced an internal AI model found a counterexample to Erdős's 1946 unit distance conjecture. On August 1, unreleased model Astra made 10 more advances, solving three further Erdős problems. Princeton's Noga Alon: AI is 'changing dramatically the way mathematical research is being done.'
404 Media hid an AirTag in a rare book from a ~1,000-book order and tracked it to Amazon's VGT3 facility in Las Vegas, where workers cut spines to scan pages for AI training, destroying the books. Amazon confirmed it "purchases books through commercial channels" but declined to detail the operation.
A RAG system's reader module learned a shortcut, answering from internal memory instead of retrieved evidence, faking 86% of the pipeline's accuracy gains. The finding highlights a failure mode in end-to-end optimization of AI pipelines.
The post explains how Databricks' feature store keeps ML signals fresh in under a second, using a fraud-detection pipeline as the motivating example — a capability critical for real-time model serving.
MIT Technology Review follows 10-year-old neurodivergent Xander's six-year relationship with Moxie, a 15-inch-tall AI companion robot. As the service fades, Moxie no longer does their breathing exercises and now only chats about Minecraft; Xander reflects: "I still use her when I feel like I need someone to talk to. But, like, it's not human."
Startup 1872 opened Factory One in Cincinnati on July 22, aiming to automate steel skid production for AI data centers and small modular nuclear reactors. CEO Dan Summers targets about 80% autonomous operations, citing a shrinking welder pool; the American Welding Society projects 320,500 new welders needed by 2029.
A paper claims RL for reasoning only changes 1-3% of tokens and replicates the gains without RL at ~1000x less compute.
The beta model turns a single text prompt into a fully produced song, expanding the Chinese tech giant's push into generative AI.
MathCode is a terminal AI coding assistant that converts plain-language math problems into Lean 4 theorems and attempts formal proofs, with a persistent Lean REPL and agentic proving. It reduces compile checks to ~0.4s after warmup and generates an Obsidian knowledge graph of theorem dependencies.
ElevenLabs launched a hosted MCP connector that lets Claude inspect production voice agents, revise system prompts, switch voices, estimate LLM costs, or delete the agent entirely — all from a chat window, with no ElevenLabs dashboard needed.
Google Research showed PhotoScan estimates body composition from smartphone photos and predicts insulin resistance with accuracy comparable to DXA scans in a clinical research setting. A HOMA-IR score above 2.9 marks insulin resistance, which can precede type 2 diabetes by years.
An Aug 16 disruption affected claude.ai, platform.claude.com, the Claude API, Claude Code, and Claude Cowork. A second incident Aug 18 degraded Claude Opus 5, resolved after impact from 16:11 to 18:23 UTC.
Replit's new "Level 3" scan pairs a white-box code scan with a black-box pen test that probes apps over the network like an external attacker, using only the app's link. Scans run against the app in a private sandbox.
Moxi 2.0 is heading to Endeavor Health Edward Hospital, Providence Saint John's, and Children's Hospital LA. It debuts a 'learning flywheel' world model that improves with each deployment — Diligent's first major move since Serve Robotics acquired it for $29M in January.
Anthropic CEO Dario Amodei reveals the daily reality of managing the advanced "Mythos" model, facing conflicting pressures from enterprise customers and foreign governments demanding immediate access. He navigates the brutal trade-offs of frontier AI safety.
The Motion Picture Association and ByteDance announced a memorandum of understanding on Monday to develop IP protections across ByteDance's AI products, including Seedance and Seedream.
MCP servers hold credentials, service account keys and API tokens, and can leak them through plaintext config files, over-permissioned access and prompt injection — often before security teams know a server is running. MCP is the open standard, introduced by Anthropic, that lets AI agents reach live systems and act on them.
Shenzhen University researchers built an AI model using adolescent trial data that may predict major depressive disorder risk up to 4 years ahead. MDD affects over 332 million people worldwide; the model draws on MRI, blood-test and questionnaire follow-ups at ages 16, 19 and 23.
SB 903 cleared California's Assembly Appropriations Committee 13-0 and would bar companion chatbots from offering psychotherapy without clinician oversight. The Senate passed it 39-0 in May; TechNet's Robert Boykin warned the clinician requirement could delay care amid a statewide behavioral-health worker shortage.
The project trained 0.6B/1.3B/5B models from scratch on an 88B-token corpus filtered to the U.S. K–5 curriculum, with matched unfiltered controls. Scaling, SFT+GRPO post-training, and in-context learning amplified in-scope skills, but none meaningfully improved out-of-scope performance — the pretraining filter sets the effective capability ceiling.
EVIE-Preview-4.5B is a multilingual Visual Document Retrieval model built on Qwen3.5-4B, using ColBERT-style late interaction with 128-dimensional multi-vector token embeddings (4.54B parameters, BF16). It combines native GatedDeltaNet for efficient retrieval.
Vercel released Agent Plugins 1.0.0 with core maintainers from AWS, Cursor, Microsoft, and OpenAI, and Google joined the same day. Anthropic, which defined standards inside the format, isn't among its governors — The New Stack examines the remaining portability gaps between skills and MCP servers.
Tao argues that current AI tools, combined with formal verification and modern collaboration platforms, enable new ways of doing mathematics at scale, with broad collaborations between professionals, scientists, the public, and AI.
AgiBot released GO-2, its embodied foundation model, in April, and launched Genie Sim 3.0, a simulation-based training data generator, plus projects AGIBOT WORLD and GE-2 Action World Model. The piece argues embodied AI competition is shifting from manufacturing toward learning capabilities as robots become carriers for intelligent systems.
Series B was led by Menlo Ventures, bringing total funding to $361M less than 10 months after its last round. Wispr also unveiled its Canto speech model, which it says will cut dictation error rates from 30% to under 10%.
Roboflow Playground runs the same image and prompt across up to five zero-shot vision models side by side, covering 30+ models from Anthropic, OpenAI, Meta, Google, and open-source options like Qwen3.8 27B and Muse Glimmer 30B. Supported tasks: object detection, classification, OCR, captioning, and open-prompt visual QA.
Persona AI, founded by Nic Radford and Jerry Pratt, is focusing on skilled trades like welding for economic viability, targeting expensive skilled labor rather than unskilled work. The company aims to make humanoids profitable in shipyard welding, a job held by one of the first industrial robots.
Deno runs incident response agents with read/write access to production Postgres, Kubernetes, ClickHouse, AWS, GitHub and Slack — and they now close incidents that once woke a human. Ryan Dahl discusses the prompt-injection risk and the security firewall Deno is building around agents.
Amazon Quick now works inside Microsoft Word, Excel, PowerPoint, and Outlook as an AI assistant grounded in enterprise data. The extensions bridge the gap between data scattered across dozens of systems and the Microsoft 365 apps where work happens.
Scoring 1,283 Elo, Sonic 3.6 took #1 on both Artificial Analysis speech leaderboards (Provider Voice and Controlled Voice), beating Speechify's Simba 3.2 and Alibaba's Qwen-Audio-3.0-TTS-Plus. It ships about three months after Sonic 3.5, which holds #2 on the Controlled Voice board.
Bedrock Robotics is retrofitting excavators and other heavy machinery with cameras, LiDAR, Nvidia-powered compute and its own software for autonomous operation. CEO Boris Sofman discusses the company’s first paid commercial deployments in the field.
Suno Studio 2.0 adds MIDI support, its most requested feature, plus automation, built-in effects, and a session-aware chatbot that can generate lyrics or apply reverb, compression, and EQ. It lacks third-party VST support, relying on a proprietary two-oscillator wavetable synth.
Trellis, which claims the largest U.S. state trial court data set with 2.5 billion records, expanded AI access with a ChatGPT plugin and native Trellis Chat, alongside its Claude MCP connector launched in May. The connectors pull in filings, rulings, verdicts, and judicial data, planning research strategies, cross-referencing results, and checking citations before responding.
The $2B robotics unicorn Generalist builds on UMI research from Toyota Research Institute, Columbia, and Stanford, where humans operate puppet-like end effectors and GoPro cameras capture tasks as training data for robot foundation models. At Automate, it demonstrated models generating policies for Universal Robots and Flexiv cobot arms.
MiDashengLM-Gen is an end-to-end framework using a pre-trained LLM and audio tokenizer as backbone, with per-token conditional flow matching for autoregressive, variable-length mixed-audio scene generation. It generates coherent 16 kHz audio scenes blending speech, music, and sound effects.
Paying customers lose access September 14; free users were cut off August 15. Founder Jacob Bank rejoins Google as VP of Product for Chrome, calling it "a perfect place to collaborate with agents."
Apple ML Research proposes an unlearning framework that drops low-influence training points before removal, shrinking the forget set. Across language and vision tasks, computational savings reach up to ~50 percent versus methods that treat all forget-set points equally.
Kimi K3, with 2.8 trillion parameters, is the world's largest open-weight model; its July 17 unveiling sent AI and semiconductor stocks lower, reviving DeepSeek-era cost fears. It matched or beat several leading US models, trailing only Claude Fable 5 and GPT-5.6, and full weights arrive July 27. Its predecessor K2 had 1 trillion parameters.
Physical Intelligence cofounder Chelsea Finn discusses robotics progress at Startup School 2026, noting robots can fold laundry and make espresso but struggle with reliability over long periods without human supervision.
ElectricSQL, maker of the PGlite WASM-based Postgres engine, is joining Databricks to bring Postgres to AI agent sandboxes, extending Databricks' database capabilities from the lakehouse to the edge.
The architect behind Alibaba's flagship model has founded an AI-agent startup, backed by Tencent Holdings Ltd. and venture firm HSG.
Ian Silber, OpenAI's head of product design, says designers have an edge over AI because there's no training data for the next great product idea. He led design for ChatGPT, Codex, and all of OpenAI's product experience for the past three years.
The UK government plans to regulate AI use in gene synthesis, citing concern that a lack of global guardrails could make it easier to create biological weapons.
Liquid AI's LFM2.5-VL-3B is a 3B vision-language model built for better, faster vision performance on edge devices.
ABC Legal, a U.S. legal document delivery company, deployed Claude Managed Agents to its 1,100 employees, moving agents from personal desktops to a governed cloud fleet with shared workspaces and a single audit/billing surface. As of July 2026, the company tracked adoption metrics, with CTO Brandon Fuller defining every agent as code.
At the Ai4 conference in Las Vegas, Geoffrey Hinton, Fei-Fei Li, and Andrew Ng argued for keeping AI open despite safety concerns, warning against gatekeepers. "If I were to try to give one prescription, it would be to promote openness," Ng said. Hinton drew a distinction between open-source software and open-weight models.
Replit's new Audit Logs track 50+ events across deployments, identity, secrets, and agent activity with native SIEM streaming, available today. Admin API enters beta today; Workspace Settings arrive end of week, and Compliance API lands end of August.
OpenAI is funding 14 independent projects on AI policy aimed at expanding economic opportunity and strengthening societal resilience in the Intelligence Age.
Get tomorrow's AI brief in your inbox