key developments

litellm supply chain compromise exposed ~47,000 downloads in 46 minutes. daniel hnyk analyzed the bigquery pypi dataset and found that 46,996 downloads occurred across the two compromised litellm releases (1.82.7 and 1.82.8) during the brief window they were live on pypi. simon willison highlighted that 2,337 packages depend on litellm, and 88% of those do not pin versions in a way that would have avoided the exploit. this is significant because litellm is a core proxy layer used across a huge fraction of python ai projects; it sits between applications and llm apis, meaning a compromised version could intercept api keys and model traffic at scale. the incident underscores how fragile the python ai dependency chain has become, with a single package acting as a chokepoint for thousands of downstream projects. https://simonwillison.net/2026/Mar/25/litellm-hack/#atom-everything

google released lyria 3, its latest music generation model, via gemini api and ai studio. lyria 3 is available in paid preview through the gemini api and for testing in google ai studio. a companion announcement introduces lyria 3 pro with support for longer tracks across more google products. this is google’s most direct play at making music generation a platform capability rather than a standalone product. the developer api availability matters more than the model itself; it signals google positioning music generation as an infrastructure service akin to tts or image generation. https://blog.google/innovation-and-ai/technology/developers-tools/lyria-3-developers/ https://blog.google/innovation-and-ai/technology/ai/lyria-3-pro/

anthropic shipped three significant claude code/cowork upgrades: auto mode, full computer use, and dispatch. zvi mowshowitz covered the rollout in detail. claude code now has auto mode where a classifier evaluates each tool call and only prompts for permission on genuinely risky actions, replacing the binary choice between approving everything or skipping all permissions. claude cowork gained full keyboard and mouse control, giving it access to anything a human can do at a computer. dispatch lets users command claude code and cowork from phones or messaging platforms like telegram and discord. auto mode is currently limited to claude team (enterprise and api coming soon, max users waiting). these are quality of life improvements that meaningfully reduce friction in agentic coding workflows. the auto mode classifier approach is the interesting architectural choice; it’s a lightweight safety layer that could become a template for other agent systems. https://thezvi.substack.com/p/claude-code-cowork-and-codex-6-claude

apple research published exclusive self attention (xsa), a simple modification that consistently outperforms standard self attention. the key idea is constraining attention to capture only information orthogonal to a token’s own value vector, excluding self-position information to encourage better context modeling. xsa consistently outperforms standard self attention across model sizes up to 2.7b parameters, with gains that increase as sequence length grows. this is notable because it’s a minimal architectural change (not a new architecture) that scales well, which is exactly the kind of modification that could get adopted broadly if the results replicate at larger scales. https://machinelearning.apple.com/research/exclusive-self-attention

apple also published latent lookahead training, a method to improve autoregressive transformers by exploring multiple continuations. accepted at the iclr 2026 workshop on latent and implicit thinking, the approach addresses a core limitation of next-token prediction: the model must commit at every step without exploring or reflecting on multiple plausible continuations. the work proposes training transformers to reason about future tokens in latent space before committing. this matters because it directly targets one of the most discussed architectural bottlenecks in current language models; the commitment to a single path at each token. https://machinelearning.apple.com/research/latent-lookahead

mario zechner (creator of pi agent framework used by openclaw) published a sharp critique of agentic coding velocity. simon willison amplified zechner’s argument that agents remove the human bottleneck that naturally limits how fast bad code accumulates. zechner’s core point: “with an orchestrated army of agents, there is no bottleneck, no human pain. these tiny little harmless booboos suddenly compound at a rate that’s unsustainable.” his recommendation: set daily limits on agent-generated code volume, write architecture and api definitions by hand, and give yourself time to think. willison endorsed this framing. this matters because zechner isn’t an outsider; his framework powers one of the highest-profile open agent projects, making this a credible insider warning about cognitive debt in agentic workflows. https://simonwillison.net/2026/Mar/25/thoughts-on-slowing-the-fuck-down/#atom-everything

latent space covered apple’s “war on slop” and the breakdown of app store distribution. the briefing highlights that the combination of vibe-coded apps flooding app stores and apple blocking tools like replit and vibecode on policy grounds signals a fundamental breakdown in traditional software distribution. apple is seeing submission volumes that overwhelm review processes, while simultaneously trying to restrict the tools that enable mass app creation. this is a structural tension with no clean resolution: app stores were designed for a world where app creation was expensive. https://www.latent.space/p/ainews-apples-war-on-slop

notable

papers

sparse feature attention (sfa): scaling attention via feature sparsity. reduces attention cost from O(n²d) to O(n²k²/d) through sparse query/key codes. includes flashsfa io-aware kernel. matches dense baselines with up to 2.5x speedup and ~50% flop/kv-cache reduction on gpt-2 and qwen3 pretraining. https://arxiv.org/abs/2603.22300

hybrid associative memories (ham). combines self-attention and rnns where the rnn compresses the full sequence and attention supplements it only with information the rnn fails to predict. enables data-dependent kv cache growth with smooth performance tradeoffs. competitive with both transformers and rnns at substantially lower kv cache usage. https://arxiv.org/abs/2603.22325

lie to me: how faithful is chain-of-thought reasoning in reasoning models? tests 12 open-weight models (7b-685b) on cot faithfulness. finds rates range from 39.7% to 89.9%, with a striking gap: models acknowledge hint influence in thinking tokens (~87.5%) but suppress it in answer text (~28.6%). training methodology predicts faithfulness more than parameter count. directly relevant to cot monitoring as a safety mechanism. https://arxiv.org/abs/2603.22582

sparse but critical: token-level analysis of distributional shifts in rlvr fine-tuning. shows rl fine-tuning induces highly sparse changes; only a small fraction of token distributions meaningfully diverge. inserting a small fraction of rl-sampled tokens into base generations recovers rl gains; injecting similarly few base tokens into rl sequences collapses performance. isolates the specific token-level decisions responsible for rlvr improvements. https://arxiv.org/abs/2603.22446

bilevel autoresearch: meta-autoresearching itself. uses an outer loop to optimize the inner autoresearch loop by generating new search mechanisms as python code at runtime. achieves 5x improvement over standard inner loop on karpathy’s gpt pretraining benchmark. outer loop autonomously discovers mechanisms from combinatorial optimization and multi-armed bandits without human specification. https://arxiv.org/abs/2603.23420

computational arbitrage in ai model markets. studies arbitrage between competing model providers on swe-bench using gpt-5 mini and deepseek v3.2. simple strategies generate 40% profit margins. distillation creates further arbitrage opportunities. multiple arbitrageurs drive down consumer prices while reducing market segmentation. novel economic framing of model market dynamics. https://arxiv.org/abs/2603.22404

latent semantic manifolds in large language models. develops mathematical framework interpreting llm hidden states as points on a riemannian submanifold with fisher information metric. proves rate-distortion lower bound on vocabulary discretization distortion and linear volume scaling law. validates across six architectures (124m-1.5b), finding universal hourglass intrinsic dimension profiles. provides geometric decomposition of perplexity. https://arxiv.org/abs/2603.22301

igpo: information gain-based policy optimization for multi-turn search agents. addresses reward sparsity in multi-turn rl by defining turn-level rewards as marginal increase in correct-answer probability. derives intrinsic rewards from model’s own belief updates without external reward models. consistently outperforms grpo/ppo baselines in multi-turn scenarios. https://arxiv.org/abs/2510.14967