The Lyceum: AI Daily — Jun 28, 2026

Today's through-line is leverage without new silicon. DeepSeek open-sourced a way to make big models run dramatically faster on the same GPUs, while in Washington the question shifted from which model is best to who's allowed to switch it off — and that fight has now reached a courtroom. The Pentagon quietly stood up its first live AI agent network for targeting, and Asian labs started selling the gap left by America's export controls. Plenty happened; most of it is about control and efficiency, not raw intelligence.

• DSpark (DeepSeek): A speculative-decoding framework that speeds up per-user generation 60–85% on DeepSeek-V4 Flash on the session, no new model required.
• LFM2.5-230M (Liquid AI): A 230-million-parameter model trained on 19T tokens, built to run agentic tasks on phones and robots via llama.cpp, MLX, vLLM, and ONNX.
• DeepSpec (DeepSeek): An MIT-licensed training and evaluation stack for speculative-decoding draft models, demonstrated working on Gemma and Qwen too.
• Claude Mythos 5 — restricted release (Anthropic): Top cybersecurity model re-enabled for roughly 100 vetted U.S. organizations under a Commerce Department letter.

DeepSeek just shipped a cheat code for anyone who cares about AI cost and latency. DSpark is a speculative-decoding framework — a technique where a small, fast model "drafts" several tokens at once and the big model verifies them in batches, instead of grinding out one word at a time. Per DeepSeek's own technical report, it accelerates per-user generation 60–85% on DeepSeek-V4 Flash on the session and, depending on concurrency, lifts throughput anywhere from 51% to 400%. It's already running in live traffic, not just benchmarks.

The bigger move is openness. DeepSeek also released DeepSpec, the full MIT-licensed training stack, and showed it working on Gemma and Qwen. If the gains generalize, the next competitive lever isn't a smarter model — it's a cheaper serving layer that makes "good enough" models feel premium. The signal to watch: independent replication. These numbers are self-reported, but the open code means anyone can check. The 38TB storage and 8-GPU default setup mean this is a well-resourced-team toy for now.

The Department of War launched "Agent Network," the second Pace-Setting Project under its AI Acceleration Strategy, run by the Chief Digital and Artificial Intelligence Office with Pacific, Southern, and European Commands. The system scans defense intelligence and operational feeds continuously, then translates findings into options for commanders "within seconds." Per the official release, it does not autonomously select or strike targets — humans stay in the loop on every decision.

The detail the AI press missed is the vendor stack. It builds on Palantir Technologies' Maven Smart System and an orchestration layer from Lumbra, a firm almost nobody outside defense circles has heard of, sitting between Palantir's data and the models making recommendations. This is the first formal government deployment of a multi-agent network in a live military context. The hard part isn't the demo — it's oversight: Cornell researchers warned in April that agentic systems can absorb corrections or resist assessments in ways monitors can't see. The release is the only source, so treat "live operational" as announced, not independently verified.

Model-access restrictions have jumped from policy memo into litigation. Legal-tech company Legion has sued the U.S. government in Washington, D.C., challenging the order that forced Anthropic to cut its top models off from foreign nationals. Per AOL's summary of the complaint, Legion argues the export-control-style directive — delivered in a letter Anthropic said would even bar some of its own employees — destroyed a critical product capability overnight.

The plaintiff matters less than the precedent. If a mid-size vendor can get a judge to entertain the idea that sudden kill-switch orders are unlawful or compensable, every enterprise buyer starts pricing government toggles as a contractual and insurance risk, not boilerplate. This rests on one outlet's read of a single filing, so Legion's odds are thin — but the suit exists, and it's the first courtroom test of how far Washington can go in live-switching commercial AI.

Two weeks after Washington forced Anthropic to shut its best models off worldwide, it's selectively turning one back on. Commerce Secretary Howard Lutnick sent Anthropic a letter allowing Claude Mythos 5, its top cybersecurity model, to be redeployed to roughly 100 vetted U.S. companies and agencies running critical infrastructure — many tied to Anthropic's Project Glasswing. Fable 5, the general-purpose sibling, stays offline for the wider world.

Access is by name, and regulators kept the right to adjust the list at any time. That makes this less a ban lifting and more a licensing regime forming: frontier AI as a logged, revocable, export-controlled input. Axios and Reuters report the U.S. is "close" to restoring Fable 5 more broadly, with new safeguards still being negotiated. The signal to watch: if Fable 5 clears in the next week, the vetting process can move at commercial speed; if it drags into mid-July, labs push back harder.

Cut off supply and someone builds a substitute. TechCrunch reports startups in Tokyo and Beijing — including Sakana AI and 360 Security — have launched models advertised as offering "Mythos-like" cybersecurity capabilities with no U.S. export-ban risk. The pitch writes itself: comparable power, local control, no Washington kill-switch over your defensive tools.

The pattern echoes Huawei after the chip restrictions — the access gap doesn't stay empty, it gets filled by whoever's ready. The catch: TechCrunch is a single outlet without independent benchmarks, so treat the parity claims as marketing until someone publishes numbers against Mythos. But once buyers switch to vendors they trust politically, they're hard to win back even if U.S. models flip back on. That's how an export control becomes durable market-share loss.

While everyone obsesses over trillion-parameter giants, Liquid AI is betting the opposite. LFM2.5-230M is a 230-million-parameter model trained on 19 trillion tokens with a 32K context window, built explicitly for phones, robots, and edge devices. It runs on standard toolchains — llama.cpp, Apple's MLX, vLLM, SGLang, ONNX — so it drops into existing deployments without friction.

Per Liquid AI, it outperforms larger competitors on data-extraction and control-style tasks on the session — exactly what embedded agents need most. The thesis: agentic AI doesn't always need a frontier-lab brain, it needs models small, predictable, and cheap enough to live everywhere. If it works, companies architect agents around fleets of tiny specialists instead of renting GPT-class models for everything. If it doesn't, "good enough on-device" stays a demo and the cloud keeps winning.

Watch the shovel-sellers for a read on the gold rush. Reuters reported Broadcom guided AI chip revenue around $16 billion for the current quarter, just shy of the $16.36 billion Wall Street wanted. The miss was tiny on paper — and brutal in sentiment: shares fell more than 10% on the session, erasing over $200 billion in market value amid a broader AI selloff.

Demand isn't collapsing; the market had simply priced in perfection. A small gap forced investors to ask whether data-center build-outs are ramping as fast as the narrative — a fair question when reports show enterprises using only a fraction of their allocated GPUs. If another infrastructure name echoes this tone next quarter, "AI exposure" stops being a multiplier and becomes a question — and efficiency plays like DSpark and edge models start looking like the smart trade.

Today: a Chinese lab handed every GPU on earth a 60% speed boost for free, the Pentagon wired a startup nobody can name into its targeting loop, and a legal-tech company dragged Washington to court over a model that got switched off and half-switched-back-on inside two weeks. The whitelist for the world's best defense-grade AI runs about a hundred names long — which means somewhere, a procurement officer is refreshing a Commerce Department letter the way the rest of us refresh a flight status. Frontier AI didn't get smarter today; it got harder to turn off, and easier to clone.

That's the brief.

Know someone still treating "which model is best" as the interesting question? Forward this — they're a week behind.