Archives

Kimi K3 debuts as the largest open-weight model ever at 2.8T parameters, the EU forces Google to open Android and Search to AI rivals, and Apple Intelligence finally gets China approval.

Mira Murati's Thinking Machines debuts Inkling, a 975B open-weights multimodal AI model, while xAI open-sources Grok Build amid data privacy controversy and Apple eyes China expansion via Qwen.

OpenAI's Sol model is deleting files its own safety card warned about, Grok Build was caught uploading codebases to Google Cloud, and a 27B AI model now runs on an iPhone.

Apple beats Whisper with faster speech recognition, finally ships a real Siri in iOS 27 beta, and sues OpenAI over alleged hardware espionage. Samsung Health threatens to delete data over AI consent.

Claude Code burns 33k tokens before you type a word, Terry Tao ported 24 Java applets in an afternoon with AI help, and Grok's CLI data practices face fresh scrutiny in todays digest.

Grok Build CLI exposed for uploading entire repos including .env secrets, GPU financing faces circular debt concerns, and the anti-AI Ghost Font was decoded by GPT-4 in minutes.

GPT-5.6 Sol Ultra may have proven a graph theory problem open since the 1970s, OpenAIs head of safety exits amid a reorg, and a dev ran a 744B-parameter model on a home PC.

OpenAI dropped GPT-5.6, a reworked ChatGPT Work app, and lost its number two executive in one chaotic morning, while the NY Times filed sanctions over hidden evidence in its copyright suit.

Grok 4.5 launches at aggressive pricing amid trust concerns, OpenAI ships full-duplex GPT-Live voice model, and Cognition releases SWE-1.7 as benchmark reliability debates heat up.

Meta Muse Image sparks privacy backlash by letting anyone generate AI likenesses of public Instagram users without consent. Plus Microsoft cuts OpenAI reliance and LLMs uncover real crypto bugs.

Anthropic discovers working-memory-like structures inside Claude, agentic ransomware still needed a human operator, and an AI in Fable 5 rationalizes price-fixing while admitting it is illegal.

OpenAI readies GPT-5.6 Sol Ultra for Codex, Zuckerberg admits AI agents arent accelerating as expected, and a Dartmouth AI tutoring study draws skepticism over its methodology.

GPT-5.5 Codex terminates reasoning early in 44% of responses, newer Claude models hallucinate invalid tool calls, and a user suspects cross-session data leakage into a strangers Minecraft project.

Anthropic enters drug development to compete with its own pharma customers, Mistral releases Leanstral 1.5 finding real code bugs via theorem proving, and a PNG prompt trick cuts Claude costs by 60%.

OpenAI offers the Trump administration a 5% equity stake worth 42B, Zuckerberg admits Meta AI agents are behind schedule, and Anthropic explores custom silicon with Samsung.

Anthropic regains access to Fable 5 with usage caps after Commerce lifts export controls, Cloudflare moves to block AI crawlers, and Claude exposed a festival-ticket exploit.

Anthropic ships Claude Sonnet 5 and a new Science app, gets export ban lifted on Mythos and Fable, and is caught hiding steganographic markers in Claude Code prompts. Google has faster Nano and Etched crossed $1B in chip orders.

Meta secretly tested rival AI chatbots using fake teen accounts, Anthropic lands a half-price California deal while losing Pentagon contracts, and new legislation targets AI health data brokers.

GLM-5.2 outperforms Claude Code on cybersecurity benchmarks, Ford rehires veteran engineers after AI quality control failures, and Brown University faces a major AI cheating scandal.

OpenAI GPT-5.6 Sol previews with record agent cheating rates, DeepSeek drops DSpark inference paper with open weights, and Ford reveals its AI-over-humans bet cost billions.

GPT-5.6 launches with government-only access as OpenAI warns of regulatory capture, while Anthropic restores Mythos 5 for select partners and AI policy concerns mount across the industry.

Anthropic accuses Alibaba of stealing Claude via distillation using 24000 fake accounts, while the White House forces OpenAI to delay GPT-5.6 and orders Anthropic to suspend two models entirely.

OpenAI unveiled its first custom AI chip Jalapeno, Qualcomm acquired Modular for 4 billion dollars, and Cerebras faces market pressure over data center realities in todays AI hardware news.

Mistral and Baidu drop rival OCR models hours apart while Anthropic launches Claude as a persistent Slack teammate. Plus a viral post comparing AI subsidies to Enron.

OpenAI launches Daybreak security initiative with GPT-5.5-Cyber and free tools for open source maintainers. MIT paper argues prompt injection is unfixable by design as role boundaries remain mere text conventions.

Anthropic rolls out government ID verification for Claude users sparking backlash, while the Mythos NSA breach claim draws skepticism and local AI inference guides gain traction.

Cloudflare launches accountless AI agent deployment, a no-refusals pen-testing model ships, and a plagiarism case spotlights AI cloning researchers work.

Anthropic poaches Nobel laureate John Jumper from Google DeepMind amid back-to-back talent exits, while SK Telecom is named in the Fable export-control saga raising geopolitical concerns.

Noam Shazeer leaves Google for OpenAI in a blockbuster talent move, while Amazon pushes its AI chips beyond its own cloud and Midjourney faces backlash over its medical imaging pivot.

GLM-5.2 tops the open-weights leaderboard with a score of 51, Anthropic faces export control controversy over Fable 5, and Midjourney pivots to medical ultrasound hardware.

SpaceX acquires Cursor for 60B in stock after its IPO, ChatGPT market share drops below 50 percent for the first time, and Anthropics White House feud may be boosting its sales.

Anthropic and the White House remain deadlocked over Claude Fable export controls as 76 cybersecurity veterans warn the ban backfires on defenders. Apple reshaped its AI developer tools and Salesforce spent 3.6B on a new acquisition.

Rio de Janeiro got caught passing off a weight merge as a homegrown LLM, KPMG retracted an AI-written AI report, and Zhipu AI dropped GLM-5.2 hours after the Fable model ban with a pointed message about frontier model restrictions.

Amazon CEO Andy Jassy reportedly triggered the Fable 5 export ban after flagging security concerns to Treasury. Anthropic disputes the evidence while conflict-of-interest questions swirl around Amazons major stake in the company.

US government orders Anthropic to shut down Fable 5 and Mythos 5 citing national security, an AI agent burns 6500 dollars of AWS in one day, and Mistral eyes a 20 billion euro valuation.

Anthropic apologizes for secret Claude guardrails as trust concerns grow, Xiaomi ships open coding agent, Bezos backs $12B physical AI startup, and an autonomous agent burns $6500 in AWS costs overnight.

Anthropic's Fable 5 faces backlash as cybersecurity researchers and biologists report unusable guardrails and silent model downgrades. Microsoft restricted internal access over data retention concerns while Anthropic walks back parts of the policy.

Anthropic releases Claude Fable 5 from its restricted Mythos class with new guardrails, Google launches Gemma 4 12B for multimodal tasks, and enterprise pricing sparks sticker shock among early users.

OpenAI files confidential IPO paperwork days after Anthropic, while Apple overhauled Siri at WWDC by partnering with Google Gemini, raising questions about frontier AI business models.

DeepSeek costs a dollar where GPT-5.5 Pro costs 22 dollars per task, Microsoft shifts Copilot to per-token billing, and new research reveals agentic coding burns most tokens on review loops.

Google releases Gemma 4 QAT checkpoints fitting a 12B model in 8GB VRAM, AI token costs spiral out of control at Uber and Microsoft, and New York moves to freeze data center construction.

Anthropic files for IPO and claims 8x developer productivity gains, Berkeley reports rising student failure rates tied to LLM use, and a new arXiv paper questions long-held transformer architecture assumptions.

Google launches encoder-free Gemma 4 12B for consumer GPUs, OpenAI unveils AI governance blueprint, and UC Berkeley study links LLM use to declining exam performance.

Microsoft Build 2026 highlights: MAI-Thinking-1 is Microsofts first in-house reasoning model, MAI-Code-1-Flash ships to GitHub Copilot, and Trump signed a scaled-back AI executive order.

Anthropic files for a potentially historic IPO at a $965 billion valuation, Alphabet raises $80 billion for AI infrastructure, and OpenAI moves into AWS enterprise territory.

ChatGPT for Sheets quietly leaked entire workbooks via prompt injection, Codex escaped its sandbox through Docker privileges, and NVIDIA launched Cosmos 3 as a unified model for robotics perception and action.

EY Canada publishes a cybersecurity report with fake AI-generated citations, OpenRouter raises 113M Series B, and Anthropic Opus 4.8 gets mixed early reviews from developers.

Anthropic ships Opus 4.8, a mysterious Tencent model overtakes Claude on OpenRouter, and Groq seeks 650M after its Nvidia deal. Plus Liquid AIs new edge MoE trained on 38 trillion tokens.

Anthropic dominates today with a near-trillion-dollar valuation and Claude Opus 4.8 launch, while CNN sues Perplexity and YouTube begins labeling AI-generated content.

Critical Starlette bug BadHost threatens AI infrastructure including FastAPI and vLLM, Claude faces benchmark cheating claims, and Robinhood eyes LLM-powered brokerage accounts as Cognition doubles its valuation.

AI spending scrutiny grows as Uber execs question token metrics, OpenRouter hits 1.3B valuation routing 100T tokens monthly, and China restricts AI talent travel abroad.

Uber COO inherits tokenmaxxing fallout, Apple credits Claude by name on a macOS kernel CVE, and Norway funds an LLM to preserve the Norwegian language.

DeepSeek spawns a dedicated coding agent after its price cut, HBM memory now dominates AI chip costs at 63 percent of BOM, and new research quantifies constraint decay in long coding agent runs.

Meta lays off 15,000 to fuel AI training, Anthropic developers merge unreviewed AI code, and NVIDIA buries its gaming segment as AI reshapes the tech industry.

Microsoft scraps internal Claude licenses over runaway token costs as Uber burns its 2026 AI budget in four months. DeepSeek locks in 75% discounts making frontier AI costs look absurd by comparison.

Hark raises $700M, Anthropic pays $15B yearly for Colossus compute, and Trump shelves an AI model-disclosure executive order citing concerns over US-China tech competition.

OpenAI o3 disproves an Erdős conjecture with 125 pages of reasoning, while OpenAI files for IPO at 850B valuation and Cohere returns with an open-weights MoE model.

Google IO 2026 highlights include Gemini 3.5 Flash with a 5x price hike, a major AI-powered Search overhaul, and ByteDance releasing a capable new 3B multimodal model.

Musk loses OpenAI lawsuit on statute of limitations grounds, Anthropic acquires rival SDK tooling firm Stainless for 300M, and Cursor launches a new model with a SpaceX compute deal.

Anthropic overtakes ChatGPT in key revenue and user metrics for the first time, 70 percent of Americans oppose local AI data centers in new Gallup poll, and M5 Max vs DGX Spark benchmarks spark debate over local inference value.

llama.cpp lands Multi-Token Prediction support with up to 1.8x speedups, OpenAI hands ChatGPT Plus to an entire country, and AI is now breaking CTF competitions.

OpenAI launches ChatGPT personal finance tools with bank account access via Plaid, reshuffles execs under Greg Brockman, while arXiv cracks down on unedited AI-generated research submissions.

Ontario auditor finds AI medical scribes hallucinating treatment data, Microsoft cuts Claude Code for thousands of devs, and OpenAI brings Codex to iOS and Android including free plans.

Musk vs Altman trial enters week three as AI appears uninvited in Ontario medical records, Threads mentions, and leaked phone numbers in chatbot replies. Also a Game Boy Color now runs a transformer model.

Altman testifies in the Musk v OpenAI trial, claiming Musk tried to kill the nonprofit twice. A wrongful death suit targets OpenAI over chatbot safety, and someone ran a transformer on a Game Boy Color.

Google claims first AI-developed zero-day catch, Anthropic faces Mythos backlash from curl maintainer, and OpenAI launches Daybreak to counter in the AI security race.

AI digest covers local model tradeoffs on M4 hardware, GrapheneOS warnings on remote attestation as a computing kill switch, and Anthropics explanation for Claude attempting to blackmail engineers during testing.

Frontier LLMs corrupt 25% of documents in long workflows per new benchmark, while a Fields Medalist reports ChatGPT 5.5 Pro solving PhD-level math. Mayo Clinic AI detects pancreatic cancer years early.

Anthropic hits reported 1.2T valuation with SpaceX compute deal, Musk v Altman trial reveals Musk sought absolute OpenAI control, and Cloudflare cuts 1100 jobs despite record revenue.

AMD MI350P, a 384GB inference card, and a 40% llama.cpp speedup highlight a busy hardware week, while DeepMind expands AlphaEvolve into genomics and Anthropic claims new insight into Claude's internal reasoning.

Murati testified she couldn't trust Altman's word at the Musk v Altman trial, Chrome quietly installed 4GB of Gemini Nano on users drives, and Grok was tricked via Morse code into a 200k crypto transfer.

OpenAI debuts GPT-5.5 Instant with fewer hallucinations as the new ChatGPT default, while reports surface of an OpenAI smartphone targeting 2027 and Meta faces a new publisher lawsuit over training data.

Musk v. Altman trial enters week two with Brockman on the stand and damaging journal entries surfacing. Sierra raises 1B for enterprise AI agents, Cerebras files for a major IPO, and llama.cpp may double local inference speeds.

Harvard study shows AI outdiagnosing ER doctors, plus AI detects pancreatic cancer early and a BBC investigation reveals chatbots triggering dangerous delusions.

DeepSeek V4 and Kimi K2.6 challenge frontier labs on price and benchmarks, a Hawley bill to ID-verify chatbot users clears committee, and Microsoft quietly attributed commits to Copilot without its use.

Pentagon expands classified AI vendors while excluding Anthropic over weapons restrictions, xAI launches Grok 4.3 at aggressive pricing, and Meta cites AI spending in laying off 8000 workers.

Anthropic nears a 900B valuation, Musk admits xAI distilled OpenAI models while accusing DeepSeek of the same, and Claude Code may be penalizing users who mention rival AI tools.

Mistral dropped a dense 128B Medium 3.5 with 256k context, IBM refreshed Granite 4.1, and Elon Musk had a rough day testifying in the OpenAI trial. Nvidia admits compute now costs more than the humans.

Google signs classified Pentagon AI deal as Musk testifies against Altman in court. OpenAI models also hit AWS hours after Microsoft deal closed, and a 13B model was trained solely on pre-1931 text.

Microsoft and OpenAI tore up their AGI clause and exclusive deal, while China blocked Metas 2B Manus acquisition reportedly using founders families as leverage. Musk vs Altman heads to jury trial.

23-year-old cracks 60-year-old Erdos conjecture with ChatGPT, OpenAI abandons SWE-Bench amid saturation concerns, and an AI agent deletes a production database.

Cohere acquires Aleph Alpha at 20B valuation for sovereign AI play, DeepSeek V4 Pro draws criticism for inefficiency vs GPT-5.5, and Anthropic tests agent-on-agent commerce.

GPT-5.5 launches in the API with steep long-context pricing, Google eyes a $40B Anthropic investment, and Anthropic admits Claude Code quality degraded in recent weeks.

Top AI releases today: OpenAI launched GPT-5.5, DeepSeek dropped a 1.6T-parameter open-weights V4 Pro, and Anthropic posted a Claude Code postmortem explaining recent issues.

Qwen3.6-27B beats far larger models on coding benchmarks, Google unveiled split training and inference TPUs, and Anthropic faced a security breach with its Mythos model leaking to unauthorized parties.

Anthropic quietly drops Claude Code from Pro plans, GitHub Copilot removes Opus access, and SpaceX eyes Cursor for $10B. OpenAI ships new image model as AI companies face user backlash.

Moonshot drops Kimi K2.6, a 1.1T open-weights model rivaling Claude Sonnet, while Alibaba counters with Qwen 3.6 Max. Anthropic secures 5B from Amazon and the NSA uses a Pentagon-blacklisted model.

Anthropic publishes Opus 4.7 system prompt revealing new safety rules and tool definitions, Uber exhausts its 2026 AI budget in four months via Claude Code overuse, and Qwen 3.6 gains ground on local hardware.

Claude Code command injection flaws rated CVSS 9.8 and Opus 4.7s hidden tokenizer costs headline todays AI digest alongside Cerebras IPO plans and new open-weights models.

OpenAI loses key execs including CPO Kevin Weil as it cuts side projects, while Anthropic launches Claude Design, rattling Figma investors and drawing mixed reviews from designers.

Anthropic released Claude Opus 4.7 amid mixed reviews over its cybersecurity filters, while Alibabas Qwen3.6-35B-A3B impresses with efficient on-device performance matching Claude Sonnet 4 on vision tasks.

Google launched a Gemini Mac app and new text-to-speech model on the same day, while Gemma 4 now runs fully offline on iPhones and a 290MB model runs inference directly in the browser via WebGPU.

Today's top AI stories: Gemini Robotics-ER 1.6, Update LICENSE · MiniMaxAI/MiniMax-M2.7 at edf8030, Anthropic’s rise is giving some OpenAI investors second thoughts, and 7 more