60 noviniek
Black Forest Labs launched FLUX 3, a multimodal model trained end-to-end across image, video (up to 20 seconds), and audio. Unlike prior pipeline solutions assembling separate models, FLUX 3 is trained jointly across all modalities simultaneously. It launches in limited access.
Etched, an AI chip startup developing specialized ASIC inference chips that operate without traditional GPUs, closed a $300M Series C at a $10.3B valuation. Prior reports indicated negotiations at up to $20B — the actual close confirmed a lower but still substantial figure.
The Cactus team post-trained Google Gemma 4 2B with a 68K-parameter probe that reads hidden states to predict p(wrong), achieving 0.814 AUROC vs. 0.549 for token entropy. Only 15-35% of queries need cloud escalation while matching frontier model performance.
STMicroelectronics lifted its 2026 AI data center revenue forecast to above $1 billion, up from 'nicely above $500 million,' driven by strong demand for 800G and 1.6T pluggable optics and silicon photonics.
Nvidia released a detailed white paper for its Vera data center CPU revealing 88 Olympus cores, 164MB L3 cache, and 1.2 TB/s bandwidth. In SPEC CPU 2026 int rate benchmarks, dual-socket Vera scores 925 vs. 898 for dual-socket AMD EPYC 9755.
TSMC committed $265B to US manufacturing, but US fabs cost 20–50% more than Taiwan, squeezing margins by 2–4%. The costs are expected to be passed on to customers — meaning higher AI chip prices broadly.
Poolside released Laguna S 2.1, a 118B MoE open-weight coding model trained from scratch in 9 weeks, free on Hugging Face — beating DeepSeek-V4-Flash and rivals 10x its size on agentic coding benchmarks.
Nativ is a free, open-source macOS application built natively on Apple MLX for running frontier open-weight models locally on Apple Silicon with no accounts, subscriptions, or cloud dependencies.
Etched, a startup building ASICs specifically optimized for transformer-based AI inference, is reportedly in discussions for a Sequoia-led funding round at valuations potentially reaching $20 billion.
Alibaba unveiled Qwen 3.8-Max-Preview at WAIC 2026 in Shanghai — a 2.4 trillion-parameter multimodal model. The company claims it ranks second only to Claude Fable 5, though no independent benchmarks have been published.
Huawei publicly debuted the Atlas 950 SuperPoD at WAIC 2026 in Shanghai — an 8,192-chip Ascend cluster claiming 6.7× the compute of Nvidia NVL144 and 15× the memory capacity. The full system delivers 1 EFLOPS at FP8 and 256TB of globally addressable memory, with Q4 2026 availability targeted.
OpenAI in Codex 0.144.6 (July 18) corrected the incorrectly advertised 372k tokens to the actual 272k. The developer community simultaneously launched a tracker documenting 35 unscheduled usage-limit resets with no published schedule.
Elon Musk announced xAI is completing training of a 2 trillion parameter model — 33% larger than current Grok 4.5. Release is targeted for August 2026, with speed and token efficiency expected to remain close to its predecessor.
Chinese startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model rivaling top US AI offerings, triggering a sharp selloff in semiconductor and AI stocks globally.
Google rolled out connected app integrations for AI Mode Search in the US, letting users build Canva designs, create YouTube Music playlists, or add Instacart grocery items without leaving the search interface.
China's Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model with a 1M context window that outranks Claude Opus 4.8 and ties with GPT-5.6 Sol in Arena benchmarks. Weights release planned for July 27.
Japan's government (METI), Noetra Corp. and NVIDIA launched the world's first national AI infrastructure — an AI factory with 13,750 Vera CPUs and 27,500 Rubin GPUs at 140 MW capacity. Jensen Huang accompanied the announcement and introduced the Cosmos 3 Edge model.
Gemini 3.5 Pro missed its third consecutive launch deadline on July 17, 2026. The fully rebuilt model still fails basic reliability standards. Google is reportedly considering a temporary stopgap release as an interim step.
Thinking Machines Lab, led by former OpenAI CTO Mira Murati and valued at $12B, released its first open-weights model Inkling — a 975B mixture-of-experts model designed for customizability, not raw benchmark performance.
TSMC beat Q2 2026 estimates driven by record AI chip demand. The company raised its full-year revenue growth outlook to over 40% and capital expenditures to $60–64B. It also announced an additional $100B commitment to US manufacturing.
Startup PrismML released Bonsai 27B, the first 27B-parameter model capable of running on iPhone 17 Pro at 11 tokens/second in a 1-bit variant weighing just 3.9 GB. Apple is reportedly evaluating the technology.
Since early June 2026, OpenAI's Codex encrypts instructions passed between the orchestrating agent and its subagents via a new 'multi-agent v2' protocol. Developers building production agent pipelines can no longer audit internal task delegation.
Nvidia has more than halved its approved Asian customer list after introducing a 'white list' requiring tighter compliance checks — including sending staff to customer data centers to verify end users — to prevent AI chips from being diverted to China.
ASML beat Q2 2026 guidance and raised its full-year 2026 revenue outlook for the second time this year to €43–45 billion (from €36–40 billion), driven by AI chip investment and accelerating DRAM orders for AI memory.
TSMC posted record Q2 2026 revenue of $39.6 billion, up 36% year-over-year. AI chips account for roughly a quarter of annual revenue, with 3nm capacity fully booked through Q1 2027.
Analysis revealed Claude Code sends nearly 5x more scaffolding tokens before processing a prompt than OpenCode — a gap that can quadruple with real-world configs.
Ploy.ai documented migrating a production UI-generation agent from Claude Opus 4.8 to GPT-5.6 Sol — 2.2x faster builds and 27% lower cost, but with unexpected behavioral pitfalls.
Comma.ai founder argues AI progress stems primarily from Moore's Law and general compute, not frontier labs overstating their contributions to justify massive fundraising.
Tokyo-based Sakana AI released Fugu, a multi-agent orchestration system packaged as an OpenAI-compatible API that achieves frontier performance by synthesizing outputs from multiple models — without relying on US-controlled models.
Irish data centers now account for 23% of national electricity consumption, driven by AI infrastructure concentration in Dublin, straining the grid and reigniting planning controversies.
Multiple Waymo robotaxis ran out of battery and required towing during July 4 gridlock in San Francisco — city supervisor Bilal Mahmood is demanding a formal inquiry into AV impact on public transit.
ByteDance launched Seedream 5.0 Pro on July 8, a direct rival to OpenAI's GPT-Image 2, featuring native 2K output, editable layer separation, sketch-based editing, and multilingual text rendering in 10+ languages.
IO Fund analyzes how Nvidia finances GPU clouds like CoreWeave and Nebius, which then buy Nvidia chips — a circular structure that inflates revenues on both sides.
An internal Meta memo reviewed by Reuters revealed plans to begin manufacturing its in-house ASIC chip Iris via TSMC in September 2026. Iris was co-developed with Broadcom and cleared bug-testing in six weeks. Meta aims to double compute from 7 GW to 14 GW and reduce NVIDIA dependence.
OpenAI launched three GPT-5.6 models and ChatGPT Work on July 9, 2026 — an agentic workplace assistant for long-running multi-step tasks, directly competing with Claude Cowork.
Artificial Analysis found Grok 4.5's hallucination rate doubled from 25% to 54% vs. Grok 4.3 even as raw accuracy improved. The model ranked 4th on the Intelligence Index but leads all models on agentic tool-use.
Developer vforno created Colibrì, a disk-streaming system that enables the 744B-parameter Z.ai GLM 5.2 Mixture-of-Experts model to run on consumer hardware with only 32GB RAM at 0.1 tokens per second.
SpaceXAI launched Grok 4.5, its first joint AI model built with Cursor, targeting long-horizon legal, finance, and software engineering tasks. Not yet available in the EU.
OpenAI officially confirmed the public global release of GPT-5.6 Sol for Thursday July 10 after the US government lifted restrictions that had limited initial access to vetted partners only.
Cognition released SWE-1.7, a coding agent trained on a Kimi K2.7 base that scores 42.3% on FrontierCode 1.1 — just behind GPT-5.5's 43.0% — at $1.97 per task running at 1,000 tokens per second via Devin.
Mistral released Robostral Navigate, an 8B model enabling autonomous robot navigation using only a single RGB camera and plain-language instructions, outperforming multi-sensor systems on R2R-CE.
Google DeepMind delayed Gemini 3.5 Pro to July 17, 2026, scrapping the existing 2.5 Pro architecture for a complete rebuild targeting mathematical reasoning and image quality. The model remains in limited Vertex AI enterprise preview.
Meta unveiled Muse, an AI image generator from Meta Superintelligence Labs, available free in the Meta AI app and integrated into Instagram Stories and WhatsApp.
Tencent Hunyuan released Hy3 on July 6: an open-source 295B MoE model with just 21B active parameters, a 256K context window, and a permissive Apache 2.0 license with no geographic restrictions.
SK Hynix launched terms for its $28.13 billion US Nasdaq offering under ticker SKHY on July 6 — the largest-ever US stock sale by a foreign company, surpassing Alibaba's 2014 IPO. Trading debut is set for July 10.
AMD's Ryzen AI Halo developer kit went on pre-order at $3,999 on July 6, built around the Ryzen AI Max+ 395 with 128 GB unified LPDDR5x memory and 126 TOPS of AI performance — designed to eliminate friction in local LLM development.
Tenstorrent CEO Jim Keller denied acquisition talks with Qualcomm despite earlier reports of an $8–10B deal, saying the company remains focused on building its own AI chip business based on RISC-V.
Anthropic is in early talks with Microsoft to run Claude inference on Microsoft's custom Maia 200 chips, which offer 30%+ better tokens per dollar. No deal has been signed yet.
Mistral CEO Arthur Mensch confirmed the company will open early access to an exciting new open-weight model in July 2026. Mistral is simultaneously in talks for a $3.5B round at a $23B valuation.
A trending GitHub issue with 107 HN points documents developer complaints about degraded output quality in GPT-5.5 Codex, with the community hypothesis pointing to reasoning-token clustering from RL post-training as the likely cause.
Flask author Armin Ronacher documents a troubling paradox: newer Claude models are less reliable at following non-standard tool call schemas, hallucinating extra fields — likely an artifact of RL training locked to Claude Code's native harness schema.
Meta AI chief Alexandr Wang announced at an internal town hall that the new 'Watermelon' model has reached GPT-5.5-comparable performance on key benchmarks — the first internal acknowledgment that Meta is closing the gap with OpenAI.
Micron Technology broke ground on a ¥1.5 trillion ($9.3B) factory expansion in Hiroshima, Japan, to manufacture high-bandwidth memory for AI processors like Nvidia GPUs, with shipments starting around summer 2028. Japan's METI is contributing up to ¥500B.
Mistral AI releases Leanstral 1.5, an open-source (Apache-2.0) Lean 4 formal proof engineering model with 119B total / 6B active parameters. It saturates miniF2F at 100% and solves 587 of 672 PutnamBench problems.
TensorWave engineers achieved 2,626 tokens/s/node running GLM-5.2 on AMD MI355X — 80% of NVIDIA B200 performance at over 2× lower cost, without any custom kernel development.
Japanese memory maker Kioxia has started shipping samples of its 10th-generation BiCS Flash for AI data center operators. 332-layer chips store 59% more data than the previous generation; Kioxia shares surged 7× this year, exceeding Toyota's market cap.
Anthropic has entered early-stage discussions with Samsung about manufacturing a custom AI chip on Samsung's 2nm process — a first step toward reducing dependence on Nvidia.
Apple released the Safari MCP Server, enabling AI agents to connect directly to Safari browser windows and autonomously inspect, debug, and test web applications using 16 built-in tools — from DOM inspection to JS evaluation and accessibility audits. Runs entirely locally with no external network calls.
The US Department of Commerce lifted an 18-day export-control restriction on Anthropic's flagship Claude Fable 5 and Mythos 5 models, restoring full global commercial access as of July 1, 2026.
Anthropic launched Claude Sonnet 5, a new frontier-tier model for coding, reasoning, and agentic tasks at introductory rates of $2/$10 per million tokens through August 31, 2026 — roughly 60% cheaper than its top Fable model.
