The Lyceum: AI Daily — May 25, 2026

Today the AI story splits into two channels that don't usually share a stage: an agent in Mountain View just solved math problems that had been open for half a century at coffee-money prices, and researchers in Singapore proved that a podcast can hijack your voice assistant without you hearing a thing. Underneath both: a quiet but structural shift in where AI's real costs and real chokepoints actually live — the memory on the chip, the power on the grid, and the cloud that hosts the model.

The most important thing about this story isn't the number — it's what the number means.

Google DeepMind's AlphaProof Nexus, a system that pairs large language models with the Lean formal proof assistant, has autonomously cracked 9 out of 353 open Erdős problems and proved 44 out of 492 open conjectures from the Online Encyclopedia of Integer Sequences. The cost per problem: a few hundred dollars. Two of the nine had been open for 56 years.

Paul Erdős spent his life posing genuinely hard problems — not textbook exercises, but open questions specialists have chipped away at for generations. What separates AlphaProof Nexus from previous AI math demos is the verification architecture: the AI proposes a proof, and Lean — a formal proof assistant — checks every logical step. If the argument doesn't hold, it gets rejected. The system layers an evolutionary search over a population of proof sketches ranked by Elo, and can call AlphaProof, DeepMind's reinforcement-learning theorem prover, as a focused subtool. Think of it as a research team where one member proposes, another critiques, and a third formally audits — running continuously at a few hundred dollars a problem. After each solve, DeepMind experts validated that the Lean statement faithfully captured the original conjecture, and results have been logged on Terence Tao's wiki of AI contributions.

The preprint isn't peer-reviewed yet. But the proofs are machine-checkable by anyone with Lean installed — a different and arguably higher bar. Watch whether the formal proof community accepts these as publishable contributions — that's the moment AI moves from impressive demo to scientific infrastructure. The most likely tell if it doesn't: silence. No follow-up papers citing the Lean files, no journal acceptances, no working mathematicians changing their workflow.

Everyone talks about GPUs as if the silicon die is the whole story. It isn't.

A new analysis from Epoch AI finds that high-bandwidth memory (HBM) now accounts for 63% of AI chip component costs, up from 52% in Q1 2024. HBM is the specialized memory stacked directly on top of the processor — the chip's working memory, holding the data the model is actively thinking about. The logic die — the thing most people picture when they hear "AI chip" — is no longer the dominant cost. The memory glued to it is.

This is structural, not a blip. Total component spending on AI chips grew from roughly $22 billion in 2024 to $52 billion in 2025, and memory alone contributed about $20 billion of that increase. SK Hynix has reported its HBM, DRAM, and NAND capacity sold out through the end of 2026. The hyperscalers are already pricing it in: Microsoft's $190 billion FY2026 capex outlook includes about $25 billion from higher component prices; Meta raised its 2026 capex range by $10 billion citing the same.

The practical implication: "compute scarcity" in 2026 is mostly a memory story. SK Hynix, Samsung, and Micron have more leverage over the AI arms race than coverage acknowledges. Nvidia gets the headlines; the memory vendors are setting the price ceiling. Watch the HBM4 ramp — if yields disappoint, every AI infrastructure timeline slips with it.

This one sounds like science fiction. It isn't.

Researchers from Zhejiang University, the National University of Singapore, and Nanyang Technological University presented AudioHijack at the IEEE Symposium on Security and Privacy in San Francisco this month. The technique subtly alters audio waveforms so humans hear normal sound but the AI system interprets hidden patterns as commands. You listen to a podcast. Your AI assistant hears something else entirely — and acts on it.

Attackers could hide malicious prompts inside music, videos, voice notes, or even live conversations uploaded to AI services. No malware. No device access. A poisoned YouTube clip becomes a delivery mechanism. The timing is pointed: voice assistants increasingly have tool-use capabilities — they can search the web, operate apps, execute commands. That last capability is the problem. As voice agents move from transcription into tool use, audio stops being content and becomes a command surface.

The security community has spent years hardening text-based prompt injection — the audio equivalent is essentially undefended. Watch whether OpenAI, Google, or Anthropic acknowledge this attack class in their voice product documentation. Silence would be telling.

Singapore's 8world reports that both Baidu Cloud and Alibaba Cloud have officially launched DeepSeek's large model on their platforms — meaning Chinese enterprises can now call DeepSeek through the two biggest domestic clouds instead of going directly to the startup.

This plugs DeepSeek's aggressively cheap models into the default purchasing channels for thousands of Chinese enterprises already standardized on Baidu or Alibaba. The hyperscalers can bundle the model into existing contracts, observability tools, and security controls. DeepSeek gets a distribution land grab without building a salesforce in every province.

Combine this with Reuters reporting that DeepSeek previewed a model tailored for Huawei chips, and the picture sharpens: China is assembling a fully domestic AI stack from chip to cloud to model. Watch how many state-owned entities standardize on "DeepSeek via Baidu/Alibaba" — that adoption curve will tell you how fast the China–U.S. pricing gap turns into a Western policy headache. The failure mode is also visible: if enterprise adoption clusters around one cloud rather than splitting cleanly, this becomes a distribution play, not a sovereignty play. [Source: 8world — Chinese (Simplified)]

36Kr reports that investors are effectively queueing outside major Chinese model companies, with DeepSeek the gravitational center. The detail that matters isn't the demand — it's that DeepSeek's willingness to accept outside capital, after two and a half years of self-funding, has reset the temperature of the whole Chinese market.

Stack that against DeepSeek's Huawei-tailored model preview and Shanghai's push to develop a "brain-like intelligent industry" — Fudan University's brain-inspired computing system has reportedly been adapted to run mainstream large language models — and a pattern emerges. The pricing war is becoming an industrial-policy war. Capital, chips, and state strategy are fusing. Watch whether DeepSeek, Qwen, Kimi, or MiniMax convert investor demand into infrastructure commitments around domestic silicon. [Source: 36Kr / Sina Finance — Chinese (Simplified)]

The invitations had gone out. Then the signing didn't happen. Axios reports anti-"doomer" feedback derailed President Trump's AI executive order last week, with draft language becoming a casualty of intra-White House infighting. But the machinery the order would have formalized is already running.

Reuters reported that Microsoft, Google, and xAI joined OpenAI and Anthropic in agreeing to give the U.S. government early access to models for security testing through the Commerce Department's evaluation setup. Separate Reuters reporting confirmed the broader oversight direction even as the executive order itself stalled.

This is bureaucratic drift becoming permanent infrastructure. Pre-release model review is arriving before the legal architecture is settled — which is historically how oversight regimes ossify. Watch whether a revised order surfaces this week. If it does, voluntary testing has crossed into something closer to gatekeeping. If it doesn't, the de facto arrangement holds without anyone having to vote on it.

British robotics company Humanoid plans to deploy up to 2,000 robots at Schaeffler plants, with an initial rollout at two German sites between December 2026 and June 2027. That date matters because it turns "someday" into an operations timeline.

Schaeffler is not a theme park. It's an industrial supplier with assembly lines and quality controls. The deal includes a separate five-year actuator supply agreement — a phrase that usually means somebody is planning for volume rather than vibes. How much of the work will be truly autonomous versus tightly scripted, and how much will land on schedule, remains to be seen. But this is how the category matures: not with one magic robot, but with procurement contracts and phased rollouts.

The milestone to watch is operational commitment, not robot charisma. Look for proof-of-output disclosures — units moved, labor hours replaced, uptime achieved. The failure mode is the rollout slipping past June 2027 with no public numbers, which would tell you the demos still haven't survived contact with a real factory floor.

Today a math agent cracked 56-year-old problems for the price of a nice dinner, a podcast can now reach into your phone and whisper commands your ears will never catch, and a German auto-parts supplier ordered two thousand humanoid coworkers like they were forklifts. The grid operator gets to switch off OpenAI's data centers when it's hot out — which is, in fairness, the most American sentence ever written about artificial intelligence.

Back tomorrow.

Forward this to the friend who still thinks the GPU is the expensive part.