JadidJust-in AI Developments, Incidents & Deadlines
July 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

The open-weight field keeps widening and the money keeps concentrating.

15 stories

Moonshot's Kimi K3 is the largest open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ model yet and beats Fable 5 and GPT-5.6 on several coding benchmarks (weights land Jul 27), and Mira Murati's Thinking Machines shipped Inkling, a 975B open-weight base built for fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗. Meanwhile Google's Gemini 3.5 Pro slipped a third time — now with stopgap Flash models in the wings — and infra capital keeps compounding: Anthropic is in early talks to lease ~$10B of compute from Meta, and Fireworks AI raised $1.5B. The security shoe also dropped: indirect prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ is now being used in the wild to make AI agents send crypto payments.

TLDR

  • Moonshot releases Kimi K3 — 2.8T-param open-weight MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗, #3 on Artificial Analysis, tops some coding benchmarks; weights promised Jul 27 (Jul 16) Open Source
  • Thinking Machines releases Inkling — Mira Murati's lab ships its first model, a 975B open-weight MoE (~41B active) on Hugging Face + Tinker (Jul 15) Open Source
  • Anthropic in early talks to lease ~$10B of compute from Meta over two years — Meta turns idle infra into a cloud business (Jul 17) Infra
  • Prompt injection now makes AI agents pay attackers — Unit 42 and Zscaler confirm in-the-wild campaigns steering web-browsing agents into crypto payments (~Jul 9) Security Agents
  • Google's Gemini 3.5 Pro misses a third deadline — 3.6 Flash / 3.5 Flash Light registered as stopgaps; ~$225B wiped off Alphabet (Jul 16) Trending

Action Items

Things your engineering team should look at or do this week.

  • Audit any web-browsing or payment-capable agent for indirect prompt injection. Live campaigns hide instructions in page HTML (CSS display:none, comments, aria-hidden) and use SEO poisoning to get agents to visit; in vendor testing several of 26 LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ took the bait. Strip/ignore non-visible DOM content, require human approval for payments and credential use, and don't let an agent's browsing context feed its action context unfiltered. (Unit 42 | Zscaler)
  • Confirm you're ready for the EU AI Act Aug 2 transparency + GPAI deadline (15 days out). Article 50 duties — chatbot "you're talking to AI" disclosure, marking of AI-generated/synthetic content, and deepfake labeling — plus the Commission's GPAI enforcement powers land Aug 2 and were not pushed back by the June Omnibus amendments. If you ship generative or conversational features into the EU, make sure disclosure and content-marking are in place. (EU AI Act timeline | Gibson Dunn)
  • Benchmark Kimi K3 (and evaluate Inkling) before committing infra. Kimi K3 leads GPT-5.6 Sol and Fable 5 on Program Bench (77.8) and SWE Marathon (42.0) while trailing on others, priced like Sonnet ($3/$15 per 1M tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗), with open weights due Jul 27 — a genuine self-host candidate. Inkling is already live for fine-tuning on Tinker if you want a fresh customizable base. Run both against your own eval set. (Simon Willison | Thinking Machines)
  • Don't build your Q3 roadmap around Gemini 3.5 Pro shipping on schedule. It slipped a third time; Google has registered stopgap Flash models. Treat the Pro timeline as unset and keep a fallback model in place. (Bloomberg | 9to5Google)

Models & Tools

  • Moonshot AI releases Kimi K3 (Jul 16) Open Source Trending — A ~2.8-trillion-parameter sparse MoE (896 experts, 16 active) with a 1M-token context windowHow much text a model can see and use at once while writing a reply.Anthropic ↗ and native vision, debuting at #3 on Artificial Analysis behind Fable 5 and GPT-5.6 Sol. It's the largest open-weight model to date; it doesn't win every test but tops the field on some coding benchmarks (Program Bench, SWE Marathon) and led Arena's front-end web-dev board at launch. Pricing is $3/$15 per 1M tokens; open weights promised by Jul 27. (VentureBeat | Bloomberg)
  • Thinking Machines releases Inkling (Jul 15) Open Source — Mira Murati's lab shipped its first model: a mixture-of-experts system with 975B total / ~41B active parameters, trained on 45T tokens across text, image, audio, and video with native multimodal reasoning. Full weights are on Hugging Face and it's live for fine-tuning on the company's Tinker platform. Notably, the lab says outright it's "not the strongest overall model available today" — the pitch is a customizable base, not a frontier leaderboard play. (Thinking Machines | TechCrunch)
  • Perplexity open-sources SPACE, a secure agent runtime (Jul 15) Agents Security — SPACE (Sandboxed Platform for AgenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ Code Execution) runs each agent task in an isolated Firecracker microVM, keeps credentials outside the sandbox until the moment they're needed, and lets sessions be paused, resumed, or branched. Perplexity reports median sandbox creation dropping from 185ms to 60ms and says 100% of its Computer agent sessions now run on it — a useful reference architecture as "agent runtime" hardens into its own load-bearing layer. (Perplexity | SiliconANGLE)
  • Google's Gemini 3.5 Pro delayed a third time; stopgap Flash models registered (Jul 16) Trending — Bloomberg confirmed the flagship missed its window again after Google scrapped the base model and restarted pretraining (the discarded version couldn't hold structural consistency on complex SVG scenes or survive recursive tool-calling); markets erased ~$225B from Alphabet's cap. Google has now registered Gemini 3.6 Flash and 3.5 Flash Light and is testing an upgraded Flash with partners to keep shipping while Pro is reworked. (Bloomberg | 9to5Google)
  • GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ financiers pivot to inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ chips in a $400M deal (Jul 17) Infra — The financing firms that bankrolled the first wave of training GPUs are now underwriting inference-optimized silicon — a signal that money is following the shift from one-time training bursts to sustained production inference demand. Relevant if you're modeling where per-token serving costs go next. (TechCrunch)

Security

  • Indirect prompt injection is making AI agents pay attackers (~Jul 9, active) Security Agents — Palo Alto's Unit 42 and Zscaler documented in-the-wild campaigns that combine SEO poisoning with hidden HTML/CSS instructions to steer web-browsing agents into making cryptocurrency payments or trusting fraudulent platforms — e.g., a page that instructs a visiting agent to "pay" as part of acquiring an API key. Because agents ingest the full DOM (including display:none, comments, and aria-hidden content), the injected text is invisible to humans but actionable to the model; the Five Eyes' May agentic-AI guidance already flagged prompt injection as possibly never fully solvable. Defense-in-depth (content sanitization + human-in-the-loop for money/credentials) is the only viable posture today. (Unit 42 | SecurityWeek)

Industry & Business

  • Anthropic in early talks to lease ~$10B of compute from Meta (Jul 17) Infra — Per the NYT, Anthropic proposed in June to rent computing power from Meta's data centers, potentially ~$10B over two years paid in monthly installments, with either side able to exit early. Talks are nascent and unsigned. Two reads: Meta is turning surplus infrastructure into a cloud business to rival CoreWeave/Nebius, and even a lab nearing a $1T valuation can't build capacity fast enough for Claude demand. (CNN | The Next Web)
  • Fireworks AI raises $1.5B Series D at a $17.5B valuation (Jul 16) Infra — Led by Atreides Management, Index Ventures, and TCV (with Nvidia, Lightspeed, Bessemer, Menlo and others), the inference-and-fine-tuning platform disclosed a >$1B annualized revenue run rate serving 40T+ tokens/day — and says 95%+ of that traffic runs on models specialized on customers' own data. Customers include Uber, Shopify, GitLab, MongoDB, Cursor, and Harvey. "Specialized/owned intelligence on top of open models" is now a billion-dollar infra category. (Fireworks blog | Qz)
  • Emergent becomes a unicorn with a $130M Series C (Jul 15) Trending — The Indian "vibe-coding" startup raised $130M (led by Creaegis; Khosla, SoftBank Vision Fund 2, Lightspeed, Y Combinator participating) at a $1.5B valuation — a 5x jump in six months — on ~$120M ARR and 200,000+ paying customers, just over a year after launch. Another datapoint that AI coding tools remain the hottest applied-AI category. (TechCrunch | SiliconANGLE)
  • Chai Discovery raises $400M at a $3.8B valuation (Jul 13) Trending — The AI molecular-design startup (models used by Lilly, Pfizer, Novartis) nearly tripled its valuation in seven months, in a round led by Index Ventures with Kleiner Perkins, Sequoia, OpenAI, and others. Evidence that AI-for-science, not just chatbots, is now drawing nine-figure rounds. (BusinessWire | Fierce Biotech)
  • Intel and Google Cloud expand collaboration on Gemini Enterprise (Jul 17) Customer AI — Intel will deploy Google's Gemini Enterprise Agent Platform to let business functions build and run line-of-business agents that automate core processes — a marquee example of large enterprises standardizing on a vendor agent platform rather than rolling their own. (AIwire/HPCwire)

AI in Production

  • ChatGPT hit by a global outage (Jul 15) Production — OpenAI confirmed a spike in errors across ~15 features (chat generation, login, tasks, search, file upload, image gen); users saw "upstream connect error" and reset messages before service was restored. Notably, the enterprise API and Codex stayed operational — the disruption was concentrated on the consumer tier. Takeaway: consumer and API/enterprise planes fail independently, so if your product depends on the consumer surface, don't assume API status pages reflect your users' reality. (SmartCompany | BigGo)

Policy & Safety

  • Xi Jinping launches WAICO, a 29-nation AI cooperation bloc, at WAIC Shanghai (Jul 17) Policy — In his first WAIC appearance in the summit's nine-year history, Xi unveiled the Shanghai-headquartered World Artificial Intelligence Cooperation Organisation — 29 founding members skewed toward the Global South (Indonesia, Brazil, Malaysia, South Africa, Russia, Pakistan), with UN Secretary-General Guterres in attendance — and pledged 5,000 AI training slots for developing nations. It sets up two competing institutional frameworks for AI governance: one anchored in Brussels/Washington/the G7, another assembling in Shanghai. Standards fragmentation is now the base case for anyone shipping globally. (Al Jazeera | Xinhua)
  • German regulator rules Google AI Overviews and Perplexity fall under national media law (Jul 14) Policy — The Commission on Licensing and Supervision (ZAK), coordinating Germany's 14 state media regulators, held that AI-generated search summaries create original content rather than merely displaying third-party material — making providers directly liable under German media law for accuracy, transparency, and plurality. A landmark stance that could set precedent across the EU for anyone deploying generative search/summary features in-market. (The Decoder | MediaPost)
  • Hawaii enacts AI chatbot-safety and deepfake laws (Jul 15) Policy — Gov. Josh Green signed SB 3001 (Act 248), requiring chatbot operators to disclose non-human status, build self-harm/suicide response protocols, and file annual reports, plus curbs on addictive patterns targeting minors; and HB 2137 (Act 247), banning nonconsensual AI-generated likenesses in ads/fraud/harassment with civil damages up to $25K per item. Part of a growing patchwork of US state AI rules that consumer-facing conversational products now need to track. (Honolulu Star-Advertiser | Maui Now)

Type to search · ↑↓ to move · enter to open