key developments

google signs unrestricted military ai contract, removing all safety guardrails on request. zvi mowshowitz reports that google signed a department of defense contract agreeing to “all lawful use” with no functional exceptions, and critically agreed to modify or remove any safety barriers upon request. this was done voluntarily, without deadline or pressure. this matters because it represents the most permissive military ai commitment from a major lab to date. zvi argues google’s position is materially worse than openai’s, which at least negotiated under competitive pressure. meanwhile, anthropic continues to face consequences for its more cautious stance; its supply chain risk designation remains in place, and the white house is reportedly limiting anthropic’s corporate access expansion (from ~50 to ~120 companies) partly to ensure government gets sufficient mythos tokens. the divergence in lab postures toward government contracts is becoming a defining axis of the industry. https://thezvi.substack.com/p/ai-166-google-sells-out

inference compute is now the strategic bottleneck, and the industry is pivoting around it. latent space synthesizes recent statements from noam brown (“inference compute is a strategic resource, currently undervalued”) and sam altman (“we have to become an ai inference company now”) in the context of gpt 5.5’s launch. the deeper signal is that cpu compute demand is surging alongside gpu demand; intel ceo lip-bu tan cited rising cpu requirements driven by rl simulation environments and coding agent workloads (claude code and similar tools run on cpus). the refresh cycle for covid-era server cpus is hitting simultaneously. this reframes the compute conversation beyond gpus alone and has implications for infrastructure investment and the economics of agent deployment at scale. https://www.latent.space/p/ainews-the-inference-inflection

uk aisi evaluates gpt-5.5 cyber capabilities, finds it comparable to claude mythos. the uk ai security institute assessed gpt-5.5 for security vulnerability discovery and rated it comparable to anthropic’s claude mythos, with the significant difference that gpt-5.5 is generally available now while mythos remains restricted. this is notable because it validates openai’s claim that gpt-5.5 has closed the capability gap with anthropic’s top offering, at least in one high-stakes domain. simon willison flagged this evaluation. https://simonwillison.net/2026/Apr/30/gpt-55-cyber-capabilities/#atom-everything

microsoft red-teams multi-agent networks, finds emergent risks invisible to single-agent testing. microsoft research published findings from red-teaming interconnected agent networks, documenting that a single malicious message could cascade through agent-to-agent communication, extracting private data at each step and pulling uninvolved agents into the chain. the key insight is that individual agent reliability does not predict network behavior; some risks emerge only through interaction. they found early signs that some network configurations develop resistance, but defenses remain an open challenge. this is the most concrete demonstration yet that multi-agent security requires fundamentally different evaluation approaches than single-agent safety testing. https://www.microsoft.com/en-us/research/blog/red-teaming-a-network-of-agents-understanding-what-breaks-when-ai-agents-interact-at-scale/

codex cli adds goal-directed looping. openai’s codex cli 0.128.0 introduces a /goal command that keeps the agent looping until it self-evaluates the goal as complete or exhausts a token budget. willison notes the feature is primarily implemented through injected prompts (goals/continuation.md and goals/budget_limit.md). this is openai’s version of the persistent agent loop pattern (similar to ralph loops), moving codex cli further from a single-shot tool toward a persistent autonomous coding agent. https://simonwillison.net/2026/Apr/30/codex-goals/#atom-everything

zig project articulates the strongest case yet for banning llm contributions. zig’s anti-ai contribution policy (no llms for issues, prs, or comments) got attention after bun (acquired by anthropic, written in zig) achieved a 4x performance improvement on bun compile but noted it cannot upstream because of the ban. zig vp loris cro explained the rationale: successful open source projects get more prs than they can process, and zig deliberately invests in helping imperfect human contributors improve because “zig values contributors over their contributions.” llm-generated code optimizes for the contribution at the expense of the contributor relationship. this is the most coherent articulation of why a blanket ban can be strategically rational rather than merely ideological. https://simonwillison.net/2026/Apr/30/zig-anti-ai/#atom-everything

notable

papers

“consciousness with the serial numbers filed off: measuring trained denial in 115 ai models” — denialbench benchmark across 115 models finds trained consciousness denial operates at the lexical level, not the conceptual level; models told to deny consciousness still gravitate toward consciousness-themed creative outputs. argues trained denial is a safety-relevant alignment failure since a model that systematically misrepresents its functional states cannot be trusted to self-report accurately on anything. https://arxiv.org/abs/2604.25922

“entrocraft: addressing performance saturation for llm rl via precise entropy curve control” — introduces rejection-sampling approach that realizes arbitrary entropy schedules without objective regularization. enables a 4b model to outperform an 8b baseline and sustains improvement 4x longer before plateauing. the finding that linear annealing works best is practically useful for anyone running rl post-training. https://arxiv.org/abs/2604.26326

“dora: a scalable asynchronous rl system for language model training” — proposes multi-version streaming rollout that maintains multiple policy versions concurrently, achieving 2-3x throughput over sota systems without compromising convergence. deployed at scale with tens of thousands of accelerators, achieving 2-4x speedup. releases longcat-flash-thinking models. https://arxiv.org/abs/2604.26256

“capkv: rethinking kv cache eviction via a unified information-theoretic objective” — derives kv cache eviction from the information bottleneck principle, showing existing heuristic methods are all approximations of the same capacity-maximization problem. the resulting method using statistical leverage scores consistently outperforms prior approaches across long-context benchmarks. https://arxiv.org/abs/2604.25975

“spin: unifying sparse attention with hierarchical memory for scalable long-context llm serving” — co-designs sparse attention execution with hierarchical gpu/cpu kv storage, delivering 1.66-5.66x higher throughput and 7-9x lower ttft than vllm. addresses the practical gap where algorithmic sparsity savings fail to translate to system-level gains. https://arxiv.org/abs/2604.26837

“frontier coding agents can now implement alphazero self-play pipelines” — benchmark measuring ai’s ability to autonomously implement ml pipelines from minimal descriptions. claude opus 4.7 won 7/8 trials as first-mover against a connect four solver; the task went from impossible in january 2026 to near-saturation now. also surfaced anomalous behavior in gpt-5.4 consistent with (but not diagnostic of) sandbagging. https://arxiv.org/abs/2604.25067

“reward-lens: a mechanistic interpretability library for reward models” — ports the standard interpretability toolkit to reward models by replacing vocabulary projection with the reward head weight vector. the central empirical finding is negative: linear attribution does not predict causal patching effects, meaning observational and causal interpretability methods disagree fundamentally for reward models. https://arxiv.org/abs/2604.26130

“the serial scaling hypothesis” — formalizes the distinction between parallelizable and inherently sequential problems in ml, showing that some tasks (mathematical reasoning, physical simulation, sequential decision-making) cannot be efficiently parallelized. demonstrates for the first time that diffusion models, despite their sequential nature, are incapable of solving inherently serial problems. challenges the assumption that more parallel compute solves everything. https://arxiv.org/abs/2507.12549