AI · Daily · May 7, 2026

thee poles, not two: claude, gpt‑5.5, hy3

Anthropic didn’t just raise limits; it locked in a sovereign‑scale cluster by partnering with SpaceX’s Colossus 1, gaining access to over 220,000 NVIDIA GPUs for Claude.[3]

That extra headroom immediately showed up as doubled 5‑hour rate limits and removed peak‑hour throttling for Claude Code’s Pro, Max, and Team plans, fixing one of users’ loudest complaints.[4]

OpenAI counter‑programmed on the quality axis instead of the hardware axis, with GPT‑5.5 Instant cutting hallucinated claims by 52.5% on high‑stakes prompts relative to its predecessor.[5]

GPT‑5.5 simultaneously leads at least one marketplace in both usage and earnings and is reported to be roughly 4–5x cheaper than Claude Mythos for comparable capability.[6]

The curveball is Tencent’s Hy3 preview model, which just processed 3.66T tokens on OpenRouter and grabbed the top leaderboard slot, turning what looked like a duopoly into a three‑pole race in actual usage.[7]

the open/local triad finally looks like a stack

On the open/local side, Qwen 3.6 27B is now effectively a coding and agent workhorse, with Multi‑Token Prediction delivering roughly 2.5x faster inference on supported setups.[8]

The same model has demonstrated context windows up to 262k tokens on 48GB GPUs, which was frontier‑only territory not long ago.[9] Benchmarks and user reports consistently show Qwen 3.6 beating Gemma 4 on coding and agentic tasks, while Gemma is preferred for planning, language nuance, and emotional tone.[10]

Gemma 4’s design—multi‑token prediction drafters plus decoupled attention in the 26B variant—lets it feel “bigger than it is,” helping models like Gemma‑4‑31B land high on code leaderboards despite struggling with tool calls and intricate coding.[11]

DeepSeek V4 rounds out the stack as a terminal‑first coding agent that many users treat as a GPT‑3.5‑class but cheaper model, with the company reportedly nearing a $45B valuation in its first fundraising round.[12]

throughput is becoming a software problem

Inference speed is starting to look like a software problem, not a hardware ceiling: the open‑source GB10 Solution Atlas engine pushes Qwen‑class 35B models past 100 tokens per second in FP8 while avoiding the PyTorch stack entirely.[13]

Multi‑Token Prediction does similar magic on mainstream stacks, giving Qwen 3.6 roughly a 2.5x decoding speedup on GPUs like the V100 when enabled.[14]

The catch is that MTP still needs around 3GB of extra VRAM headroom and doesn’t reduce core model compute, so it’s great for wall‑clock latency but less of a FLOPs savings hack than people assume.[15] vLLM 0.20.0 arrived with day‑0 MTP support for Gemma 4 and turnkey Docker images, while AMD’s MI355x on SGLang reportedly delivered more than a 10x throughput jump per GPU since launch.[16]

On the training side, NVIDIA and Unsloth’s recipe of packed‑sequence metadata caching and better MoE routing is good for about a 25% LLM fine‑tuning speedup, which compounds with all these inference tricks.[17]

agents are mostly rag with a gpu addiction

The agent ecosystem looks enormous at first glance: LangChain has already crossed 1B downloads despite being only a few years old.[18] Major clouds have converged on AG‑UI as a shared frontend for agents, and LangSmith now tracks around 300M agent runs per month at Clay.[19]

Under the hood, research using CrewAI and LangGraph shows these agents often burn far more compute than simple chatbots for relatively modest gains, which is why tools like Shadow now exist just to regression‑test their behavior.[20]

A lot of what’s marketed as “autonomous agents” is effectively RAG with extra ceremony—benchmarks and practitioner writeups bluntly describe many agents as glorified retrieval wrappers around LLMs.[21]

Teams are discovering observability the hard way, retrofitting logging and metrics after agent incidents, hence projects like MetaLens for Metabase‑based debugging and the broader push to treat observability, drift, and performance as first‑class design constraints instead of afterthoughts.[22]

platform creep and security go from vibes to exploits

The browser is now an LLM runtime whether you asked for it or not: Chrome has been caught silently dropping a roughly 4GB Gemini Nano model onto users’ machines, triggering EU legal questions about consent and data use.[23]

On the “local is safer” side, Ollama just shipped a critical unauthenticated memory‑leak bug nicknamed Bleeding Llama, which can expose sensitive data from local LLM sessions if unpatched.[24]

Developer tools aren’t better: VS Code’s Copilot integration started auto‑adding itself as a commit co‑author, while users report growing distrust of Microsoft’s Copilot and Windows 11 changes in general.[25]

At the application layer, people are accidentally leaking API keys through tools like Cursor and Copilot, while persistent‑memory agents have been shown vulnerable to deliberate memory‑poisoning attacks and data exfiltration.[26]

Even core model training has security and ethics overhangs, with lawsuits alleging large‑scale scraping of protected books to train models like Meta’s, and users expressing fresh concern about how much control they really have over where their data lands.[27]