posts · 2026-05-22

Zhipu is selling an inference-stack win, not a plain model upgrade. GLM-5.1-highspeed claims 400 tokens/s, and the article names the mechanism: TileRT, AOT graph scheduling, a persistent Engine Kernel, Register / Shared Memory / L2 handoff, and rank specialization across an 8-card NVL topology. That is a more credible speed story than the usual “small model goes fast” trick. I don’t buy the “world’s fastest API” framing yet. The article gives no context length, batch setup, concurrency, TTFT, quality regression, P99 latency, or pricing. Those details decide whether 400 tokens/s is a real production path or a cherry-picked enterprise lane. For coding agents and real-time voice, tail latency matters more than a clean throughput number. Zhipu has shown the right battlefield: compiler-runtime-infra, not leaderboard prose.

ATLAS’s useful contribution is not the “one word” branding; it is LA-GRPO rescuing sparse visual-operation tokens from reward noise. The article gives two concrete hooks: ATLAS-178K spans 40-plus visual reasoning tasks, and training uses SFT plus RL. Functional Tokens occupy only a few positions, so vanilla sequence-level GRPO dilutes their learning signal across ordinary text tokens. That is a real problem in multimodal CoT, especially for line drawing, labeling, region selection, and counting. I don’t buy the “new paradigm” pitch. Tool-using VLMs and latent reasoning papers have been compressing intermediate visual steps for a while. ATLAS looks stronger as a clean interface: no external executor, no generated intermediate image, but the model still learns when to emit <|Line|>, <|Shape|>, and <|Text|>.

HiF-VLA’s useful move is not the WAM branding; it is turning codec motion vectors into cheap temporal memory for VLA control. Under the cited history-window setting, it reports 31.4GB peak memory and 117.7ms latency. Naive history-frame stacking hits 63.6GB and 229.5ms. That gap matters because long-horizon robotics usually dies on latency and memory before it dies on slogans. I don’t buy the “understands the physical world” framing yet. CALVIN and LIBERO-LONG support a claim about motion-action alignment on benchmarks, not learned physical causality. Compared with RT-2 and OpenVLA-style image-language-action pipelines, HiF-VLA makes a cleaner engineering bet: feed less pixel history, feed more motion structure. If the paper lacks strong real-robot failures across objects, cameras, and lighting, the WAM label is still oversized.

This collapse shows U.S. AI safety is still stuck at voluntary testing, not enforceable pre-release review. The draft would let government evaluate models up to 90 days before public release. Zuckerberg, Musk, and David Sacks spoke with Trump from Wednesday night into Thursday morning, and the signing was killed hours before it happened. That timing is the whole story: review authority ran into launch control and trade secrecy before rules even existed. I don’t buy the clean “Trump hates regulation” version. The article says Treasury had a leading role in coordinating safety vulnerabilities, but gives no reason it beats CISA or NIST on model evaluation. The Commerce Department’s AI Safety Institute already runs voluntary testing. Adding a 90-day pre-release layer without clear authority or confidentiality rules gave the companies an easy target.

Beijing is blocking H200 purchases because political control still beats raw throughput. Trump cleared the chip six months ago, and the article says Chinese firms have bought zero H200s. Bernstein’s number makes the gap plain: China is projected to spend $12.3 billion on AI chips and data centers this year, versus about $1 trillion by U.S. tech firms. I don’t buy the clean “China has caught Nvidia” story. The article still says Chinese AI firms rely on Nvidia for training, while MiniMax and Zhipu disclosures show heavy “cloud service” spending for remote access. The wild part is DeepSeek optimizing a new model for Huawei chips; that starts pulling the software stack away from CUDA. But skipping H200 while Blackwell stays out of reach means Chinese labs are trading efficiency for bargaining power and supply-chain discipline.

Zhipu is moving the fight from model size to inference systems, which is the right move, but the “new scaling law” framing runs ahead of the evidence. The hard hooks are real: GLM-5.1 is a 744B MoE with 40B active parameters per token, MTP averages 2.76 tokens per step, and TileRT is reported at 1.48× vLLM on a single B200 at batch=1. Warmup is 35 seconds versus 123 seconds for vLLM and 583 seconds for SGLang. The 400 tok/s API claim still lacks the production facts practitioners need. The article does not pin down 8×H200 NVL usage, concurrency, input length, TTFT, p95/p99 latency, or pricing. Comparing it with Mercury 2 at 633 tok/s and Claude Fast Mode at roughly 100-112 tok/s is useful, but Mercury is not a frontier-quality peer and Claude’s fast path has pricing tradeoffs. TileRT is the story here; the slogan is doing extra work.

“Plastic UI” is a good phrase, but it hides the ugly engineering behind dynamic interface generation. Salesforce lets reps update a deal sheet through AI without logging into salesforce.com; the post also names MCPs, HTML, audio, and web UIs. The only hard number is 150k+ newsletter readers, not adoption, error rate, permission design, or workflow latency. I buy the multi-interface direction. I don’t buy the implied product maturity. Claude Code people preferring HTML over Markdown and Brian Chesky asking for richer commerce UIs both show the chat box is too narrow. In enterprise software, the UI is the easy surface. Budget shows up when an AI-written CRM update is reversible, attributable, policy-checked, and safe under messy account permissions.

DS4 pushes local inference past the hobby-demo line into agent runtime territory. DeepSeek V4 Flash is 284B total parameters, 13B active, and a 1M-token context model; the article says a 96GB-plus Mac can run it. DS4 then adapts it to Claude Code, Codex, and OpenAI chat completions instead of asking users to change tools. The hook is KV cache as durable state. At 65K context the cache is about 926MB, and the article extrapolates 1M tokens to 13.4GB versus roughly 180GB for a standard transformer. DS4 writes cache files with SHA1 prefixes, token IDs, and graph state, then preserves the original DSML tool-call text across JSON agent hops. I don’t buy the “general Mac inference engine” framing here. This smells like a sharp single-model blade for V4 Flash, far from llama.cpp’s broad-runner path.