The Lyceum: AI Daily - Jul 05, 2026

It was a holiday weekend in the U.S., and the model-drop machine went quiet — no major releases from any of the big labs in the past 24 hours. But the interesting story right now isn't a new model anyway. It's the plumbing underneath them: who gets the chips, who's learning to build without them, and what it actually costs to run the agents everyone keeps promising. China's chip situation has quietly split into two contradictory tracks, and the numbers coming out of the agent economy don't flatter the sales pitch.

No major model or tool launched in the past 24 hours — the July 4th weekend thinned out the release calendar. For context, here's what shipped earlier this week (all more than 24 hours old):

• Claude Sonnet 5 (Anthropic): Described by Anthropic as its "most agentic" Sonnet-series model, built for long-running agent workflows and tool use.
• Gemini 3.5 Flash (Google): First model in the Gemini 3.5 family, optimized for complex agentic workflows and coding.
• Nemotron 3 Nano Omni (NVIDIA): A ~30B omni-modal mixture-of-experts model unifying text, vision, and audio, claiming up to 9× higher throughput than comparable open multimodal models — per NVIDIA's announcement.
• DeepSeek-V4-Pro (DeepSeek): Mixture-of-experts model with a 1-million-token context window, aimed at long-running agentic tasks.
• Devstral 2 (Mistral): A 123B model optimized for coding agents, released quietly alongside a smaller 24B variant.

The most important thing about Meituan — China's answer to DoorDash — is no longer that it delivers dumplings. On June 30, it released LongCat-2.0, a 1.6-trillion-parameter language model it claims is the first Chinese frontier model trained — not just run — entirely on domestic chips. (Parameters roughly measure a model's complexity; 1.6 trillion puts it in the same weight class as the biggest models on Earth.)

"Trained on" matters more than the size. Flagship models from DeepSeek and Zhipu can already run inference — the act of answering a question — on Huawei-made chips, according to The Star. But training a model is a different beast: running one is like driving a car, training one is like building the engine from scratch. Meituan says it did the whole build on a 50,000-card domestic cluster using AI ASIC superpods, with hardware reportedly tied to Huawei's Collective Communication Library and Atlas-950 SuperPods, per Nation Press.

Meituan claims LongCat-2.0 performs comparably to Google's Gemini 3.1 Pro. That's a company claim, not an independent benchmark — treat it as a starting point. But if the training claim survives scrutiny, it's the strongest proof yet that China's chip ecosystem can sustain frontier development without Nvidia, which shortens the runway on U.S. export controls considerably. The signal to watch: whether independent researchers can verify the training logs in the coming days. If they can't — or find Nvidia hardware in the stack — the whole narrative deflates.

Two DeepSeek chip stories landed earlier this year, and together they describe a policy Washington still can't fully digest.

Track one: back in late January, Reuters reported that Beijing conditionally approved DeepSeek to buy Nvidia's H200 chips — Nvidia's second-most-powerful, one tier below the fully-banned Blackwell. ByteDance, Alibaba, and Tencent reportedly got clearance for 400,000 H200s between them. Nvidia CEO Jensen Huang told reporters he'd received no actual orders yet and believed China was still finalizing licenses.

Track two: separately, in February, Reuters reported — and a senior Trump administration official confirmed — that DeepSeek had already trained a frontier model on smuggled Blackwell chips, likely clustered at its Inner Mongolia data center. So the U.S. is simultaneously negotiating to sell China its second-best chips while confirming China obtained its best ones illegally. White House AI Czar David Sacks and Huang argue that shipping legal chips discourages Huawei from catching up; China hawks fear diversion to military use. What changes if this holds: the export-control framework gets partially unwound by negotiation rather than legislation. The observable signal — Commerce Department enforcement action, or its conspicuous absence.

Everyone's excited about agents. A study surfacing in Korean tech press this weekend puts a number on the cost that deployment announcements quietly skip: AI agents use up to 136.5 times more energy per question than simple Q&A chatbots.

In plain terms: asking a chatbot a question costs roughly one unit of energy. Asking an agent to do something — research, write and run code, manage a workflow — can cost 136 units. Energy is the primary driver of inference cost, and inference cost is what decides whether an AI product survives at scale. The enterprise pitch right now is essentially "replace human workflows with agents." At two orders of magnitude more compute per task, the unit economics look very different from the demos. The winners here will be whoever cracks inference efficiency — quietly one of the most important races in AI, even if it earns fewer headlines than model launches. Watch for energy to show up as a line item, not a footnote, on the next round of enterprise earnings calls. [Source: Maeil Business Newspaper — Korean]

Most legal AI tools have a dirty secret: they're great at summarizing documents they've already seen and terrible at navigating large, unfamiliar document sets the way a junior associate would. On Saturday, LlamaIndex published legal-kb, a public reference application on GitHub built on its Index v2 platform.

The clever part is the toolkit. It gives an agent four distinct ways to work a corpus: retrieve (semantic search), find (keyword and citation lookup), read (full document access), and grep (pattern matching across the whole set). Real legal research isn't one operation — it's a sequence of different search strategies depending on what you're hunting for. This is a reference blueprint, not a product a firm deploys tomorrow. But the design pattern — arming agents with several specialized retrieval tools rather than one generic search — is the right architecture for any domain full of dense, structured documents. The signal: whether enterprise legal vendors adopt or fork this pattern over the next quarter. If they do, LlamaIndex just set the foundation for a category.

Most industrial knowledge doesn't live in text — it lives in engineering drawings, CAD files, and schematics that current AI models mostly can't parse. Shanghai AI Lab announced a large model for visual reasoning and decision-making that can automatically produce engineering drawings and build 3D assemblies.

If it works as described, that would let AI systems participate directly in manufacturing design — not just the text-based paperwork around it. The claims come from Chinese-language coverage and haven't been independently verified in English-language outlets, so treat the capabilities as preliminary. But the direction matters: physical AI — systems that interact with the designed and built world, not just the written one — is one of the field's most consequential frontiers, and it's moving faster inside China's industrial AI ecosystem than most Western coverage reflects. What to watch: whether the lab publishes a paper that lets outsiders evaluate the claims. [Source: The Paper — Chinese]

A dumpling-delivery company casually training a trillion-parameter model on homegrown silicon, DeepSeek smuggling banned chips into a data center in Inner Mongolia while politely applying to buy the legal ones, and an AI agent quietly incinerating 136 times the electricity just to answer the same question a chatbot handles for pennies. The export-control map keeps getting redrawn — but the people actually paying attention aren't lobbying Washington, they're on GitHub, benchmarking open models on AMD hardware and wiring up browser agents while the lawyers argue over which Nvidia chip is contraband. Quiet weekend; loud subtext.

Forward this to the friend who still thinks the chip war is about chips.