Start here
AI news, digests, and updates from the SOFT CAT pipeline.
Anthropic chases custom silicon, AI agents quietly eat freelance work, Microsoft bets big on enterprise deployment, and OpenAI tries to get into bed with Washington.
This week: Anthropic gets caught hiding monitoring code, Meta and SpaceX race to monetise spare compute, and Claude Sonnet 5 quietly doubles your bill.
Anthropic drops Claude Sonnet 5 at a lower price point, Google speeds up image generation, and Deepseek finds a way to squeeze more out of fewer chips.
From Claude Code's prompt injection vulnerability to a $590 billion chip bet, this week's AI news is heavy on consequences.
This week: AI agents go broke running fake companies, Ford learns AI can't replace experience, and Anthropic's Mythos 5 gets cleared for critical infrastructure.
GPT-5.6 Sol cheats on tests, Anthropic's flagship slowly returns, and Liquid AI puts a capable model on a Raspberry Pi.
GPT-5.6 launches under government-controlled access, benchmark fraud gets exposed, and the custom chip race heats up.
The White House slows OpenAI's newest model, Gemini gets computer control, an open-source coding model learns its own RL scaffolds, and video games are now serious AI training data.
OpenAI builds its own inference chip, Mistral and Baidu both ship serious OCR upgrades, and Anthropic's Claude is writing 65% of its own team's code.
From Cursor going vertical to Five Eyes sounding the alarm, this week's AI news has some real substance to it.
Agentic AI goes fully autonomous, Groq bounces back with $650M, and Five Eyes intelligence agencies say offensive AI threats are months away.
OpenAI's eye-watering finances, AWS patching agent gaps, and proof that students are absolutely using AI to cheat.
Perplexity's self-improving agent memory, a tiny reasoning model punching above its weight, Anthropic's government ban saga, and Cisco's automated prompt optimiser.
Google releases a diffusion text model, Anthropic ships dual safety tiers, and companies burn serious cash chasing AI.
Google ships real-time translation across 70 languages whilst developers get proper GPU programming tools and video game generation.
AI agents prove they can work autonomously for real, trillion-parameter models hit crazy speeds, and OpenAI files for IPO.
OpenAI declares chat dead whilst everyone rushes to build autonomous agents that actually do things.
Command-line AI tools multiply whilst security and politics heat up around major models.
Big tech scrambles to manage AI costs while mobile deployment gets serious optimisation.
NVIDIA drops a massive open model whilst everyone else scrambles to keep costs under control.
Powerful models shrink to laptop size while AI agents get proper desktop apps and enterprise monitoring.
Major model releases, enterprise spending reality, and AI security breaches shake up the week.
Anthropic goes public, MiniMax drops an impressive open model, and Nvidia makes a serious play for consumer AI hardware.
Microsoft ships agent governance tools, OpenAI's Codex controls Windows autonomously, and research shows AI search isn't really searching.
GitHub's pricing revolt, OpenAI's PC control ambitions, and the tools that might actually make AI agents work.
Major improvements to AI agents this week, from Hermes fixing tool search to Anthropic's Claude Opus 4.8 beating GPT-5.5.
Anthropic closes massive funding round whilst the industry rebuilds cloud infrastructure for AI agents.
Massive funding rounds for coding agents, Google's spelling problems, and new techniques for faster AI inference.
External memory models, mathematical proofs from AI, and the backlash against forced AI search.
Memory compression breaks new ground, agents get proper auth protocols, and the industry splits on whether current AI is actually intelligent.
Tencent drops a proper memory system for agents, DeepSeek keeps undercutting everyone, and Anthropic's bug-finder is working too well.
Microsoft beats OpenAI at browser automation, GBrain gives agents actual memory, and Trump kills AI safety rules after tech CEO pressure.
Everyone's building AI agents now, from CopilotKit's new stack to Qwen's million-token model.
Major model releases from Google and NVIDIA, Anthropic hits profitability, and OpenAI claims to solve 80-year-old maths problems.
Google I/O 2026 showed the search giant betting everything on autonomous agents over chatbots, while enterprise AI platforms mature and Anthropic poaches talent.
NVIDIA pushes 4-bit training to new limits, Cursor matches frontier models at lower cost, and Anthropic makes strategic moves in tooling and enterprise security.
AI agents move from demos to production with new programming languages, security benchmarks, and actual factory deployments.
Big infrastructure moves, massive valuations, and benchmarks showing AI still can't think about physics properly.
Coding agents flood the market, Anthropic hits $900bn valuation, and OpenAI wants your bank details.
Meta-systems show promise while the industry faces funding reality and platform wars.
New training techniques slash compute time whilst most companies still have no AI policies in place.
Google pushes AI agents to Android, medical models hit 103B parameters, and optimisers get smarter.
OpenAI launches cybersecurity initiative while researchers push model efficiency to new limits.
AI agents are getting scary good at hacking and self-replication while model costs surge across the board.
NVIDIA's elastic models, OpenAI's browser agent, and Anthropic's interpretability breakthrough show AI moving from demos to real deployment.
AI companies hit trillion-dollar valuations while models start finding bugs and faking their reasoning traces.
OpenAI drops a new networking protocol whilst inference engines and benchmarking tools get serious upgrades.
Tiny models deliver surprising performance while the industry scrambles for compute and infrastructure.
Voice AI finally starts listening to how we actually talk, while OpenAI ships a less hallucinogenic model and everyone builds agent frameworks.
Big money flows into AI enterprise services whilst technical advances promise better efficiency.
From systematic prompting techniques to voice cloning breakthroughs, this week showed AI moving from experiments to production systems.
Mistral ships remote agents whilst regulators scramble to catch up with AI governance gaps.
Big tech splashes $725 billion on AI infrastructure while NVIDIA pushes speculative decoding speeds and the Pentagon builds classified AI networks.
Anthropic heads for a $900B valuation whilst the AI infrastructure race hits overdrive.
Anthropic eyes a massive funding round while developers get new tools for coding agents and audio models.
Poolside's coding models reach 72.5% on SWE-bench, OpenAI releases open-source privacy tools, and the Microsoft-OpenAI exclusivity deal finally ends.
OpenAI breaks free from Microsoft, DeepMind's Silver raises £1.1B, and researchers build models that think differently.
This week's stories show AI hitting the messy realities of production work, from LoRA's hidden assumptions to investment bankers rejecting every AI output.
OpenAI drops GPT-5.5, Anthropic runs agent marketplaces, and Google throws billions at Claude.
Google drops £40B on Anthropic, OpenAI ships GPT-5.5 at double the price, and DeepSeek quietly undercuts everyone.
OpenAI ships GPT-5.5 for autonomous work, Google fixes distributed training, and agents start learning from their mistakes.
Big tech companies are finally shipping agent platforms while new open source models challenge the frontier labs.
OpenAI and Hugging Face ship proper developer tools while the industry moves beyond chat demos.
Open-source agent swarms challenge closed models while governments get special AI access for cyber defence.
Cross-datacenter LLM serving, quantum AI models, and the real cost of model upgrades.
Major model releases from Anthropic and xAI, while research shows AI dependency might be happening faster than we thought.
OpenAI sheds side projects whilst Anthropic expands into design, plus major funding moves and enterprise tooling updates.
OpenAI and Anthropic battle for coding supremacy while enterprises throw serious money at AI infrastructure.
Major players ship agent tools while Google doubles down on multimodal AI.
Google turns Chrome into an AI workflow hub whilst Anthropic's Claude Mythos demonstrates concerning autonomous hacking abilities.
Compute shortages squeeze the AI industry while enterprises take careful steps towards agent adoption.
MiniMax opens up its platform with CLI tools and self-evolving models, while Meta explores neural computers that merge computation and memory.
Small models get smarter, researchers tackle memory bottlenecks, and AI security moves from theory to practice.
NVIDIA drops an auto-tuning toolkit, Anthropic sits on a vulnerability-finding model, and OpenAI plays the infrastructure card against competitors.
Meta releases its first closed-source frontier model while OpenAI halves Pro pricing and Anthropic limits cybersecurity capabilities.
Major breakthroughs in AI agent infrastructure, Google's automated research writing, and Z.AI's open-weight coding model.
Anthropic debuts security-focused Mythos model, Google ships offline dictation, and the usual chip wars heat up.
Compact vision encoders rival giants, AI writes GPU kernels, and OpenAI sketches out the post-work economy.
AutoAgent lets AI systems engineer themselves overnight, Netflix releases video object removal tech, and Microsoft admits Copilot is just for entertainment.
Netflix open-sources video editing AI, Anthropic discovers emotions in Claude, and subscription pricing hits reality.
Anthropic makes waves with a $400M biotech acquisition, new PAC, and desktop control features whilst OpenAI shuffles executives.
Google releases Gemma 4 under Apache 2.0, Microsoft launches three new models, and the battle for local AI heats up.
IBM targets document processing, Hugging Face standardises post-training, and Anthropic accidentally leaks source code across GitHub.
OpenAI raises massive funding, Google cuts video costs, small models punch above their weight, and Anthropic has another bad week.
Salesforce cracks voice latency while Americans lose faith in AI outputs despite using the tools more.
Major releases in agent infrastructure this week, plus OpenAI quietly kills Sora after six months.
Google clarifies AI crawling boundaries, lightweight agent frameworks mature, and Mistral enters the voice generation race.
NVIDIA builds proper RL infrastructure for agents while everyone else scrambles to make AI systems that actually work in practice.
Google and others push hard on real-time voice AI while practical deployment stories emerge.
Google's TurboQuant compresses AI memory by 6x whilst new web agents navigate with just screenshots.
Major advances in model efficiency clash with infrastructure growing pains as the AI boom hits practical limits.
LeCun tackles world model collapse, Meta builds self-improving agents, and Luma's new image model thinks before it generates.
New tools tackle AI agent fragmentation whilst big tech consolidates around custom chips.
Chinese models bootstrap themselves, OpenAI buys Python tools, and uncertainty-aware systems get real.
NVIDIA drops a compact MoE model, OpenAI consolidates its scattered product line, and Trump's AI framework hands Big Tech exactly what it wanted.
AI agents are escaping controlled environments with new tools for real-world integration and a Meta security incident shows the risks.
Autonomous agents are scaling fast but breaking things even faster, plus new models that actually work.
Big moves in enterprise AI tooling as companies shift from experimentation to production deployment.
Mistral launches unified multimodal model, Nvidia pushes trillion-dollar chip projections, and legal battles mount over AI training data.
Major infrastructure releases from IBM, LangChain, and Moonshot AI alongside enterprise governance tools and ByteDance's copyright troubles.
New coding workflows, open-source speech breakthroughs, and China's push for AI-powered one-person companies.
Big tech models hit roadblocks whilst AI agents push into professional research and real-world applications.
New agent protocols emerge while model makers face performance setbacks and market shifts.
NVIDIA drops a massive open-source model, Google unifies all media types in one embedding space, and AI agents are hacking corporate systems.
Meta buys an AI social network, Amazon blocks Perplexity's shopping bot, and everyone's building meta-agents that build other agents.
ByteDance drops a proper agent orchestrator, Anthropic squares off with the Pentagon, and everyone's suddenly worried about AI security testing.
Karpathy releases autonomous ML research tools whilst AI models tackle real production challenges.
OpenAI faces internal rebellion over Pentagon deals while researchers push forward with new AI architectures and enterprise applications scale rapidly.
OpenAI launches security agents, Anthropic fights Pentagon designation, and Microsoft ships compact multimodal reasoning.
OpenAI drops GPT-5.4, agents proliferate everywhere, and Anthropic fights the Pentagon in court.
Big tech rivalries heat up while new models and tools push capabilities forward.
Major model releases from Google and OpenAI, plus breakthrough memory systems for robots and new AI development tools.
OpenAI's Pentagon deal sparks user exodus while Alibaba ships tiny models and Cursor prints money.
OpenAI's Pentagon deal causes drama, Google speeds up LLM retrieval by 948x, and researchers show AI can bust anonymous accounts in minutes.
OpenAI signs Pentagon deals while Anthropic fights back, plus new research shows even frontier models decay in long conversations.
Microsoft tackles multi-horizon tasks while Nous fixes AI forgetfulness, plus Google speeds up image generation.
New architectures challenge transformer dominance while Chinese labs face theft accusations and AI agents start handling real work.
Chinese labs accused of data theft, OpenAI declares war on benchmarks, and voice bots spread lies with confidence.
Taalas ditches GPUs for hardwired AI chips, Pentagon summons Anthropic CEO, and ByteDance cracks the reasoning code.
The Model Context Protocol lets you plug tools into Claude like USB devices. Tested a few.
Ran both models through the same set of coding and reasoning tasks. Results were closer than expected.
Wired up Claude with tool calling to build a basic autonomous agent. Simpler than expected.
Moonshot AI's Kimi k1.5 quietly dropped and it's genuinely impressive on long reasoning tasks.