Wolf Digest — Monday, May 11, 2026

A million-hour curated video corpus focused exclusively on human activity, with a pretraining recipe that beats prior open VLMs on Kinetics-derived action understanding benchmarks. The scale is roughly 10× the prior open record for human-centric video pretraining; 33 upvotes on HF Daily Papers.

Cerebras Systems is expected to price its IPO Thursday at a $35 billion valuation, the largest U.S. pure-AI-silicon offering since Nvidia's secondary in the late 1990s. The Information notes that demand has been strong enough for the company to consider raising the price range — Bloomberg reported the same on Friday. If the offering follows the CoreWeave pattern (priced at $40, closed Friday at $114 a share), Cerebras enters the public markets as a credible alternative compute story to Nvidia despite continuing to burn through cash. The CS-3 wafer-scale architecture has accumulated a meaningful customer book on inference workloads where its on-chip SRAM advantage matters most, and the company has been increasingly visible on Artificial Analysis's leaderboards for high-token-rate gpt-oss serving.

A sparse-attention scheme that learns a mixture of fixed-pattern indexers (block-sparse, strided, BigBird-style) rather than a single sparsity prior, plus a routing head that picks per-layer. Reports near-dense quality on 128k-context retrieval-augmented eval at a fraction of the FLOPs.

A new multitask benchmark for code search that decomposes the task into retrieval, ranking, and rewrite components, plus a baseline model that sets a new SOTA on the union of evals. The framing is explicitly against generic embedding retrieval, arguing that production code-search needs first-class ranking and rewrite stages. 22 HF upvotes.

An on-policy distillation recipe for flow-matching generative models that closes the inference-step gap to single-step student samples without the FID degradation typically seen in consistency-distillation baselines. Demonstrated on text-to-image and class-conditional generation.

An RL recipe for multimodal search agents that runs many parallel visual queries and learns to drop low-utility branches early. Reported wall-clock savings on multimodal retrieval benchmarks with no accuracy loss.

An efficiency-oriented successor to the original Byte Latent Transformer that closes the wall-clock gap to BPE-tokenizer baselines while keeping the byte-level vocabulary. Aimed squarely at low-resource-language and multilingual deployments where tokenizer-bias is most painful.

A method for explicit reasoning over 4D (3D + time) representations rather than collapsed video frames. Targets robotics-style spatial reasoning where temporal coherence matters; reports gains on dynamic-scene QA benchmarks.

A controllable, interactive red-teaming platform tuned for agentic models (tool use, browser use, computer use). Aims to fill the gap between unit-test-style harm evals and full-environment red-team exercises; ships with a structured adversary library.

Frames the state-tracking limitations of recurrent and SSM models as an error-control problem and introduces a new layer that explicitly bounds drift. A useful theoretical-plus-empirical contribution to the long-running parity-with-attention debate.

Proposes an anisotropic-projection objective for cross-modal alignment that beats CLIP-style contrastive objectives on zero-shot retrieval at matched compute. Suggests symmetric cosine objectives leave room on the table when the modality-specific feature distributions are themselves anisotropic.

A continuous-latent diffusion approach to language modeling that drops the discrete denoising step in favor of a learned latent space. Bridges between LLaDA-style discrete diffusion and continuous image-diffusion machinery; quality competitive with same-size autoregressive baselines on standard language-modeling benchmarks.

A framework that unifies several self-distillation variants under a single objective and shows how to interpolate between them. Practical contribution is a single recipe that recovers the gains of task-specific distillation methods (e.g., on-policy vs. off-policy, hard vs. soft labels) without having to pick one upfront.

A benchmark for evaluating agentic search systems that interleave text, image, and structured queries. Designed to expose failures of unimodal benchmarks at characterizing real-world multimodal agent stacks.

A long-context prefill scheduler that dynamically blocks inputs and reorders computation across the prefill timeline. Reports 1.4–2× speedup over chunked-prefill baselines on 64–128k contexts at matched memory.

A parallelization scheme for the stepwise-momentum variant of delta linear attention that recovers training-time throughput parity with softmax attention. Continues the slow march of linear-attention variants toward production viability.

The Information's analysis of 100 public tech companies' March-quarter earnings calls finds the AI-productivity narrative splitting in two: companies like Spotify, Uber, and Airbnb credit AI for margin improvement via flat or declining headcount, while others say rising AI infrastructure costs are compressing margins. The split is consistent with the late-2025 thesis that the gains accrue asymmetrically to high-marginal-cost service businesses, while integrated platforms eat the inference tab. Either way the 'AI is unambiguously margin-accretive' framing of 2024 is no longer the consensus.

A family of generative models that combines normalizing-flow likelihood tractability with trajectory-based parameterization. Aims to retain the exact-likelihood property of flows while matching diffusion's sample quality at moderate compute.

A two-tier MoE routing scheme — coarse task-cluster routing and fine within-cluster expert routing — that scales continual learning to 300+ tasks without catastrophic forgetting. Reports minimal interference across the task suite.

An adaptive entropy-regularization schedule for RL training of multi-turn agents. Modulates the exploration coefficient based on per-turn confidence signals; reported wins on the tau²-Bench and Terminal-Bench suites versus fixed-entropy baselines.

Investigates what makes a latent manifold easy for diffusion to model and proposes a prior-aligned autoencoder loss that yields cleaner FID/CLIP-score curves at every compute budget. Useful for anyone training a new latent-diffusion model from scratch.

A speculative-decoding variant that drafts whole blocks (rather than single tokens) and dynamically rebalances the draft tree. Continues to push speculative decoding closer to its theoretical ceiling on long-context generation.

Splits long-context inference into a shallow-network prefill and a deep-network decode, exploiting the empirical observation that prefill saturates early-layer features. Reports meaningful TTFT improvements at matched output quality.

The Information's weekend column reads the xAI–Anthropic compute deal as the leading edge of a broader pattern: frontier labs treating rivalry and supply relationships as separately optimized. Picks up the OpenAI-Google TPU deal, Anthropic-Akamai, Anthropic-Blackstone-HF-Goldman enterprise JV from earlier in the month, and now xAI-Colossus. The framing is that strategic exclusion is too expensive at frontier scale; everyone is going to be a supplier and customer to everyone else.

Combines autoregressive frame generation with diffusion-based denoising plus an agentic planner that issues per-segment objectives. Reports the strongest long-form consistency numbers to date on minute-scale video generation.

An agent architecture that decomposes a complex task into a structured plan with skill-level conditional gating. Beats flat-prompt LLM-agents on long-horizon tau²-Bench Telecom tasks.

Bridges autoregressive language modeling with normalizing-flow image generation under a single training objective. Aims at unified multimodal understanding + generation without separate decoders.

A real-time neural audio codec with asymmetric encoder/decoder design — heavy encoder, light decoder — for live streaming. Reported quality at parity with EnCodec at fraction of the decoder FLOPs.

Uses the understanding pathway of unified multimodal models to supervise the generation pathway, improving generation-side instruction following without retraining the visual encoder.

A taxonomic survey of agent-memory architectures, organized by the storage-to-experience continuum. Useful as a reading list for anyone designing a long-running agent's memory layer.

A theoretical paper proposing that the marginal price of cognitive labor will converge to a compute-anchored equilibrium as agents scale. Of interest as one of the first peer-reviewed framings of the labor-economics-of-AI question in standard micro language.

A multi-domain benchmark for evaluating language-model intent recognition, with 50K human-curated examples spanning customer support, code intent, and instruction interpretation.

A continual-adaptation scheme that builds a case library of deployment-time interactions and selectively replays them. Aimed at the gap between training-time fine-tuning and production drift.

Empirical evidence that the decision regions of modern image classifiers are simply connected with high probability — a result with implications for adversarial-robustness analysis and the theoretical understanding of network expressivity.

A deep-unfolding architecture that learns a common-PCA decomposition across source domains for downstream classification. Continues the small-but-active deep-unfolding sub-thread.

Diarmuid Gill and Liva Ralaivola of Criteo join Nathan Labenz to unpack modern adtech AI: millisecond-latency recommendation, realtime bidding, deep-learning embeddings, foundation-model integration, and what Criteo's new OpenAI partnership for product discovery in ChatGPT means for the search-engine-as-product-funnel pattern. Notable as one of the first detailed technical breakdowns of how a top-tier adtech stack is integrating foundation models.

A multi-agent system for CNC manufacturability checking, built at the AMD developer hackathon. The blog post is a worked example of an industrial-agent stack running on the MI300X — useful as a concrete pattern for anyone deploying agentic systems to the manufacturing side.

Wispr Flow reports accelerated growth in India after its Hinglish rollout, despite the well-known difficulty of voice AI in a code-switching market. Useful data point for anyone building multilingual voice products in low-resource-language settings.

Jensen Huang delivered Carnegie Mellon's 128th commencement address, framing the moment as the start of the AI revolution and drawing the parallel to his own career start at the dawn of the PC era. Light on hard news but the rhetorical positioning is worth tracking — Nvidia has been deliberately shifting from 'GPU company' to 'platform-for-the-AI-era' messaging.

A reader-facing glossary of recently-emerged AI terminology — hallucinations, RAG, RLHF, etc. Of limited interest to ML-literate readers; worth noting as a sign that the mainstream tech press is starting to standardize vocabulary.

A short read on the ergonomic and acoustic implications of voice-first workflows in shared office spaces. More cultural-trends commentary than product news; worth a line as voice AI consumes more white-collar workflow.