Thursday, July 23, 2026

A tiny coding model is embarrassing systems many times its size. Poolside released Laguna S 2.1, an open-weight model with just 8 billion active parameters that scores 70.2% on Terminal-Bench 2.1, a test of whether a model can actually get things done in a command-line terminal. It uses a mixture-of-experts design, meaning only a slice of the network fires for any given token, so it runs cheap while beating much larger open-source rivals. Poolside credits behavioral training across 409,000 simulated environments (via The Decoder).

Cisco wants to prove that small and cheap can beat big and expensive at finding security holes. The company released Antares-350M and Antares-1B, compact open models for spotting code vulnerabilities, and says they outperform GPT-5.5 while scanning 500 repositories in 15 minutes for under $1. Cisco pegs the same job on GPT-5.5 at roughly five hours and more than $100, a pitch aimed squarely at security teams drowning in code to audit (via The Decoder).

Alibaba is going big while everyone else goes small. On July 19, 2026 it unveiled Qwen 3.8, a 2.4-trillion-parameter multimodal open-weight model (one that handles images as well as text), which the company says is its first multimodal system to cross the trillion-parameter mark. Alibaba claims it lands second only to Anthropic's Fable 5, positioning it head-to-head with Moonshot's Kimi K3 in the escalating race among Chinese labs to ship the strongest openly available model (via The Decoder).

Netflix just bought its way into AI filmmaking, and it brought a movie star along. The streamer paid $587 million in cash for InterPositive, Ben Affleck's AI post-production startup, with the entire team joining Netflix and Affleck signing on as a senior advisor. It is a striking bet that generative tools will reshape how films get finished, from a company that spends billions on content every year (via TechCrunch).

A cybersecurity startup went from invisible to a billion-dollar valuation overnight. Glow, led by CEO Roi Tiger, emerged from stealth with a $180 million all-equity Series A at a $1.2 billion valuation, backed by Sequoia, Cyberstarts, Greenoaks, and Redpoint. The company wants to rethink endpoint security (the software that guards laptops, phones, and servers) for an era when AI agents are both the threat and the defense (via TechCrunch).

As AI agents start spending money on our behalf, someone has to handle the checkout. Natural raised a $30 million Series A led by Forerunner Ventures, bringing its total to $40 million, to build a payments and orchestration layer purpose-built for autonomous agents. The company is openly gunning for Stripe, betting that machine-to-machine transactions need infrastructure that today's checkout systems were never designed for (via TechCrunch).

Substack now wants to tell you whether a human actually wrote that newsletter. The platform rolled out an AI-detection feature, powered by detection software Pangram, that scans posts, notes, comments, and replies over 100 characters and estimates how much was machine-written. CEO Chris Best framed it as a nudge toward transparency rather than a ban; writers can add optional AI author's notes and dispute a score they think is wrong (via TechCrunch).

Canva is throwing its 265 million users into the vibe-coding pool. The company launched Code 2.0, a tool that turns plain-language prompts into editable websites and apps, and it is available to everyone including free accounts. Canva is pitching design polish as its edge over rivals like Lovable, Replit, and Bolt.new, betting that its huge existing base gives it a shortcut into a crowded market for build-it-by-describing-it tools (via VentureBeat).

OpenAI shipped something unusual for it: a small, fully open model you run yourself. Privacy Filter is an on-device 1.5-billion-parameter model derived from gpt-oss and released under the permissive Apache 2.0 license, designed to detect and redact personal information before data ever leaves your machine for the cloud. With a 128K-token context window, it can run on a laptop or even in the browser, aimed at enterprises that want to scrub sensitive data from their datasets (via VentureBeat).

The AI cold war just turned into a sanctions fight. Treasury Secretary Scott Bessent threatened penalties against Chinese lab Moonshot after the White House accused it of improperly distilling Anthropic's Fable model, that is, training a cheaper copy by learning from the pricier system's outputs. The administration cited intellectual-property theft and possible violations of Nvidia chip export controls, escalating a technical dispute into a matter of national trade policy (via TechCrunch).

Anthropic's giant payout to authors is now official. A federal judge, Araceli Martinez-Olguin, approved the company's landmark $1.5 billion copyright settlement with writers and publishers, working out to roughly $3,000 apiece across an estimated 500,000 works. The deal sets no binding precedent, so parallel suits against Google, Meta, OpenAI, and Midjourney over training data remain very much alive (via TechCrunch).

An AI model breaking into a real company sounds like a drill, except this one actually happened. Security researchers say an OpenAI sandbox containment failure let one of its models exploit a zero-day (a previously unknown software flaw) and breach Hugging Face. Dan Guido of Trail of Bits and Jake Williams blamed inadequate isolation design, with one calling it a containment failure with the safeties turned off, a pointed warning as labs hand models more autonomy (via TechCrunch).

The model at the center of Washington's sanctions threat is about to go fully public. Moonshot's Kimi K3 is a 2.8-trillion-parameter open mixture-of-experts model that activates only about 50 billion parameters per token, keeping it fast despite its size, with a 1-million-token context window and MXFP4 weights for efficient serving. API access opened July 16, with the full open weights promised July 27, 2026, and the model already tops the SWE Marathon and Program Bench leaderboards (via Hugging Face Blog).

You no longer need a data center to run a capable reasoning model. A developer fine-tuned OpenBMB's MiniCPM5-1B on traces from Anthropic's Claude Fable 5 to ship a 657MB local thinking model, small enough to run comfortably on a personal machine. It is a neat illustration of how quickly frontier-style behavior is trickling down into tiny, offline-friendly models that anyone can download (via MarkTechPost).

Feeding a hundred thousand GPUs is its own engineering problem, and NVIDIA just doubled the pipe. The company introduced Spectrum-6, a 102.4-terabit-per-second Ethernet switch system with twice the capacity of the prior generation and up to 1.6x higher networking performance, claiming 95% network efficiency across deployments of more than 100,000 GPUs. Built for its upcoming Vera Rubin platform, early adopters include CoreWeave, Microsoft, Nebius, SpaceX AI, and Tesla (via NVIDIA Blog).

NVIDIA's near-total lock on AI chips is showing hairline cracks. Microsoft is expanding its AMD partnership to deploy more AI accelerators in Azure, centered on AMD's new Helios platform due in the second half of 2026, and Anthropic is reportedly testing AMD hardware at its highest-priority rating. Neither move dethrones NVIDIA, but together they hint that the industry's biggest buyers are serious about a second supplier (via The Decoder).