Daily 2026 · 05 · 28
Economic Tech

Cognition raises $1B at $26B as Devin's run-rate jumps thirteenfold in a year

Lux Capital, General Catalyst and 8VC co-lead a round that more than doubles the AI coding startup's September valuation. Cognition says annualized revenue has gone from $37M to $492M in twelve months, with Goldman Sachs, NASA, Mercedes-Benz and Santander on the customer list.

Bloomberg · 2026-05-27 · Source ↗ ·

Cognition closed more than $1 billion in fresh capital at a $25 billion pre-money valuation, putting the post-money mark at $26 billion. Lux Capital, General Catalyst and 8VC co-led the round, with Ribbit Capital, Atreides Management and Founders Fund joining and existing backers Elad Gil, Soma Capital and Omri Casspi following on. The deal more than doubles the $10.2 billion figure Cognition carried after its September 2025 round.

Bloomberg reports Cognition's annualized revenue run-rate has moved from $37 million in May 2025 to $492 million today, a 13× jump in twelve months. Cognition's pitch to investors leaned on enterprise usage growth that has compounded at 50% month-over-month for the past six months. The named customer roster on the round materials includes Goldman Sachs, Mercedes-Benz, NASA, Santander and several U.S. government buyers.

Devin, post-Windsurf

The flagship is Devin, the autonomous software-engineering agent Cognition unveiled in March 2024. The product runs in a sandboxed workstation, reads tickets, opens pull requests, and ships code with limited human review. After Google acqui-hired most of Windsurf's leadership team in July 2025, Cognition bought the remaining business and brought the Windsurf IDE under the Cognition umbrella, fusing an agent-first engineering tool with an IDE that already had paying enterprise seats.

That combination is the spine of the new round. Devin handles long-horizon work — multi-file refactors, dependency upgrades, ticket-to-PR cycles. The IDE that used to be Windsurf has become the surface developers actually live in. Cognition's pitch is that the agent and the editor are now one product, and the run-rate is the proof.

Competitive frame

The round lands during a brutal price war for the AI coding layer. Anthropic's Claude Code 2.1.152 shipped the same day Cognition announced the round, with a one-keystroke /code-review --fix that mirrors the agent workflow Cognition built into Devin. OpenAI's Codex update on May 27 added local conversation search, stronger MCP and CLI profile handling, and clearer tool schemas. Google's Jules sits one tier down, and xAI's grok-build-0.1 became broadly available to SuperGrok and X Premium+ users via Kilo Code the same week.

Simon Willison, writing the same evening, argued that the broader story is that Anthropic and OpenAI both crossed into "real product-market fit" because of coding tools, with enterprise pricing shifted from flat-rate seats to API-metered tokens in April. Cognition's $1B round is the venture market's version of the same thesis: the dollars are in coding agents, the agents are eating tokens, and the buyers will pay.

"Coding agents really did change everything. These are tools which burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals."— Simon Willison, on the May 27 enterprise-pricing shift at Anthropic and OpenAI

What's not in the price

The round closed before Cognition disclosed a path to profitability. The company is burning aggressively on compute, paying Anthropic, OpenAI and self-hosted inference for every Devin run, and the disclosed run-rate is bookings rather than recognized revenue under a strict definition. Reporting in TechCrunch noted that Cognition's enterprise contracts include large multi-year minimums that pull future demand into the run-rate calculation. Investors are paying for a 50:1 multiple on annualized revenue and an even steeper multiple on cash receipts.

The other open question is durability. Anthropic and OpenAI host their own coding agents and have a structural cost advantage if the buyers begin to consolidate on a single foundation lab. Cognition's bet is that a coding-specific company can build distribution and workflow lock-in faster than the labs can build sales motions, which is the same bet Cursor was making before it raised at a $46B valuation, and the bet Replit and Vercel are making at lower valuations.

Big picture

Cognition's $1B round at $26B set the dollar mark for May 27, but the deeper signal is that the coding-agent stack is now where the money lands and where the technical work clusters. Anthropic, OpenAI and xAI all shipped on the same Wednesday: Claude Code 2.1.152 turned code review into a one-command apply, the Codex CLI got local-history search and tighter MCP handling, and grok-build-0.1 went wide for SuperGrok subscribers through Kilo Code.

The research wing pushed in parallel. Hugging Face's daily papers led with NVIDIA's LocateAnything for fast vision-language grounding, KAIST's geometry-aware 3D reconstruction, and Tencent's EvalVerse for cinematic video benchmarks. ArXiv carried MUSE-Autoskill on self-evolving agents, MobileGym on parallelized GUI-agent training, MobileMoE on on-device experts, and FinHarness, a safety wrapper for finance LLM agents that lands in the same week U.S. and U.K. regulators sharpened their AI guidance.

Underneath: Tensormesh closed $20M from AMD Ventures and CoreWeave to push KV-cache reuse as an inference-cost story, Fireworks AI is in talks at $15B, and RevEng.AI raised $15M to verify binaries shipped by other people's coding agents. Voxel51 published a survey of 700 physical-AI builders saying 74% of the industry is underinvested while GenAI sucks up budget, and the SQLite project quietly told the world it does not accept agentic pull requests at all.

More today

27 stories · 2026-05-27 — 2026-05-28
Tech

Codex adds local conversation search and tighter MCP handling

OpenAI · 2026-05-27 · Source ↗ ·

OpenAI shipped a Codex CLI update with case-insensitive search across local conversation history, stronger MCP and CLI profile handling, and richer extension and hook context. Goal mode is now generally available.

OpenAI's May 27 Codex release moves the CLI closer to the everyday-tool feel that Claude Code has been chasing for months. Local conversation history is now searchable with case-insensitive content matches and inline previews, so a developer can go back to a half-finished refactor from last week without piping JSON through grep. The release notes also flag stronger MCP and CLI profile handling, which is what users had been asking for since the MCP tunnels Anthropic shipped in May made cross-tool sessions practical.

The bigger user-facing change is Goal mode shipping as generally available. Goal mode lets the user state an outcome ("ship a deployable PR that passes CI") and lets Codex plan and execute multi-step tool calls toward it, with checkpoints. The browser-use path also got "more reliable" treatment per the release notes, alongside locked computer use for eligible Mac users and Mac app Appshots. On Windows, the TUI rendering is sharper.

Tool schemas got the most attention from the release-engineering side. Schemas are "more reliable" and remote sessions hold their state through transient failures. The docs and packaging cleanups suggest OpenAI is moving Codex out of the early-adopter posture and into something it expects to support at scale.

Why it matters

Coding agents are now the load-bearing demand driver for inference. Cognition's $1B round on the same day priced that demand at $26B. Anthropic, Google, xAI and OpenAI are all racing to make the CLI feel like the editor the developer already trusts. Local conversation search and reliable tool schemas are unsexy plumbing, but they are the difference between a Codex session that survives a crashed terminal and one that doesn't.

TechApplication

ChatGPT Enterprise/EDU adds Skills governance, upload scanning and compliance logs

OpenAI · 2026-05-27 · Source ↗ ·

An admin Skills page now lets owners review, transfer or delete workspace skills, while uploads are scanned and risky files are blocked. Compliance Logs Platform supports listing, exporting and deleting skill records.

OpenAI rolled out a governance surface for Skills inside ChatGPT Enterprise and Edu workspaces. Admins now get a dedicated Skills page where they can review workspace skills, update access, transfer ownership, and delete skills the team no longer needs. New permission toggles cover who can use skills, who can upload files, who can share, and who can publish or install skills for others.

The release also turned on automated upload scanning. Third-party skill files go through a review pass; uploads that look risky are flagged for manual inspection, and confirmed malicious files are blocked. The Compliance Logs Platform now supports listing, exporting and deleting skill records and identifies the specific skill that ran inside a conversation stream.

Skills remains in early access and is disabled by default for new Enterprise/Edu workspaces. The release reads like preparation for general availability: the IT-buyer machinery — audit, permissions, file scanning — is in place before broad rollout.

Pattern

The Skills surface is where ChatGPT becomes an integration platform rather than a chat product. The May 27 update implies that OpenAI expects Enterprise customers to start uploading proprietary skills against internal data sources at meaningful volume, and the security review pipeline has to exist before that wave hits.

Tech

Claude Code 2.1.152 ships /code-review --fix and skill access control

Anthropic · 2026-05-27 · Source ↗ ·

A one-command review flow now applies fixes directly to the working tree. Skills and slash commands can set disallowed-tools, and a new MessageDisplay hook lets users transform assistant output.

Anthropic shipped Claude Code 2.1.152 on May 27 with a feature developers had been requesting for months: /code-review --fix runs the standard review and then applies the suggestions to the working tree in one shot. The change makes the review-then-apply loop a single keystroke and moves Claude Code closer to the agent-style flow that Devin and Cursor have been building toward.

The release also tightens skill and tool governance. Skills and slash commands can now set disallowed-tools, so a piece of shared scaffolding can lock down what the model can call. A new /reload-skills command rescans the skill directory at runtime. SessionStart hooks can return reloadSkills: true and set session titles. A new MessageDisplay hook lets users transform assistant messages before they render in the terminal.

Organizations get managed plugin suggestion settings, fallback model support when the primary model is unavailable, and an Auto mode that no longer requires consent each turn. Vim mode picked up reverse history search via /. The /usage breakdown now scans session files for a more accurate per-category split.

"Fixed terminal styling degrading in very long sessions by recycling the renderer's style pool."— Claude Code 2.1.152 release notes

The bigger arc

Anthropic's developer platform announcements from earlier in May — self-hosted sandboxes for Managed Agents on Cloudflare, Daytona, Modal and Vercel, plus MCP tunnels in research preview — point at a coherent story: agents should run in places customers already trust. The 2.1.152 release fills in the local-machine corner of that story by making the Claude Code CLI feel less like a chat window and more like a programmable shell.

Tech

xAI opens grok-build-0.1 to SuperGrok and X Premium+ users through Kilo Code

xAI · 2026-05-27 · Source ↗ ·

xAI's coding-specialized model is now available to subscribers via the open-source Kilo Code IDE and CLI, with no extra signup. The model is positioned for autonomous multi-step engineering work.

xAI broadened access to grok-build-0.1 on May 27, letting SuperGrok and X Premium+ subscribers use the model inside Kilo Code, the open-source agentic coding platform. The integration runs from the IDE or the CLI with no separate signup beyond the existing subscription, putting xAI's coding-specialized model directly into a tool many developers already use as a Cursor or Continue alternative.

grok-build-0.1 was announced in early May as a model "built specifically for agentic software engineering" — it plans, writes, refactors and iterates on code across multi-step workflows. The May 27 expansion is a distribution story: rather than wait for everyone to install yet another CLI, xAI plugged the model into Kilo Code's pre-existing surface and let the subscription model do the rest. Broader rollout to subscribers happened between May 24 and 27, with the company's official confirmation on X on the 27th.

Read-out

The coding model lineup looks like four labs racing to be the default code-writing tool for the enterprise developer. Anthropic has Claude Code; OpenAI has Codex; Google has Jules; xAI now has grok-build-0.1 with first-class IDE integration. The pricing model is becoming the key differentiator: xAI is rolling the model into a flat-rate consumer subscription, while Anthropic and OpenAI switched their enterprise plans to API-metered pricing in April.

TechTrend

SQLite tells AI agents: bug reports OK, code submissions not

Simon Willison · 2026-05-27 · Source ↗ ·

SQLite published an AGENTS.md that bans agentic code contributions outright, spun up a separate forum for AI-generated bug reports, and removed the "currently" qualifier from its no-agentic-PRs policy.

SQLite, the most-deployed database in software history, has staked out a policy position that most projects have been ducking. The project's new AGENTS.md file, surfaced by Simon Willison on May 27, says SQLite does not accept agentic code. The team strengthened the language by removing the "(currently)" qualifier in a recent commit, signaling this is not a temporary stance.

The team will accept AI-generated bug reports with reproducible test cases and demonstration patches, but not for direct merge — the patches are for illustration only. The flood of AI-generated reports has been heavy enough that SQLite created a separate bug forum to handle them, where D. Richard Hipp has been working through the queue and making commits.

"SQLite does not accept agentic code."— SQLite AGENTS.md

The wider tension

The same week, Cognition raised $1B at $26B on the back of its agentic engineering platform, and Anthropic shipped /code-review --fix for Claude Code. SQLite's posture stakes the opposite end of the same debate: the project's reliability rests on a tiny set of humans who own every line, and an unbounded inflow of LLM-shaped contributions threatens the audit chain. Other infrastructure-critical projects — curl has gone further, Linux kernel maintainers have been increasingly explicit about wanting authored attribution — are testing similar policies. The AGENTS.md spec started as a way to help agents work in repos. SQLite has flipped it into a sign on the door.

Tech

Voxel51 survey: 92% of physical-AI builders say world models are the next frontier

Voxel51 / VMblog · 2026-05-27 · Source ↗ ·

A survey of more than 700 visual and physical AI practitioners finds 78% seeing ROI, 74% calling the field underinvested, and 100% reporting model failures in deployment. Annotation remains the operational pain point.

Voxel51's inaugural State of Visual and Physical AI survey, released May 27, polled more than 700 professionals working on visual, spatial, and physical AI systems. The headline finding is a sharp split between perceived value and actual investment: 78% of respondents say they are getting value from their visual and physical AI work and expect that value to grow over three years, while 74% say the industry is underinvested relative to where the opportunity sits — and the most-cited reason is that GenAI and LLMs are absorbing budget and attention.

Deployment is brittle. Only a third of respondents report that their visual or physical AI deployments have reached production. Every respondent — 100% — has experienced underperforming models in deployment, and 85% say model failures are common in their organization. The survey's authors traced the failure modes back to data: annotation pipelines, dataset quality, and the gap between annotated training data and what actually arrives in production.

Annotation in particular came out as one of the deepest pain points. Only a minority of respondents have annotated more than half of their available datasets, and 36% reported that less than half of their annotated data ever makes it into production training runs.

What's next

The forward-looking finding: 92% of respondents say world models and spatial intelligence are the next frontier for visual AI. The same group is bullish on synthetic data, edge deployment, and tighter loops between annotation and model retraining. The survey lands as NVIDIA, Meta, and a wave of robotics startups push physical-AI tooling, and as Cosmos and Isaac GR00T-style world-model platforms try to claim the spatial-intelligence layer.

Tech

LocateAnything: NVIDIA paper claims fast parallel-decoded grounding

arXiv / NVIDIA · 2026-05-27 · Source ↗ ·

NVIDIA's vision-language grounding method decodes object boxes in parallel and topped Hugging Face Daily Papers on May 27 with 108 upvotes. The model targets latency-sensitive applications like AV perception.

LocateAnything, NVIDIA's new vision-language grounding model, took the top slot on Hugging Face's Daily Papers list on May 27 with 108 upvotes. The paper's premise: most VLG models decode bounding boxes sequentially, which limits throughput and makes them poor fits for autonomous driving, robotics or live video pipelines. LocateAnything proposes a parallel box-decoding architecture that issues all candidate boxes in one decoder pass, with cross-attention between boxes during decoding to resolve overlap.

The paper claims competitive quality with previous grounding SOTA at a fraction of the latency, which lines up with NVIDIA's broader physical-AI thesis — its Cosmos and Isaac GR00T families want grounding to be a primitive that other agents call hundreds of times per scene, not a heavyweight pass. NVIDIA Research is publishing the model alongside the paper.

Read alongside

The same day's HF Daily Papers list carried Tencent's EvalVerse for cinematic video benchmarking (76 upvotes), Ropedia's SpatialBench for spatial foundation models (75 upvotes), an 11-author paper on MobileGym for verifiable mobile-GUI-agent training (56 upvotes), and KAIST's geometry-aware representation denoising for multi-view 3D reconstruction (35 upvotes). The cluster makes the same case in three different ways: the next frontier of evaluation is spatial, multimodal and grounded in the real world, not text.

Tech

MUSE-Autoskill: self-evolving agents that create and manage their own skills

arXiv · 2026-05-27 · Source ↗ ·

A Huawei-led paper proposes a four-loop architecture for agents that propose new skills, store and prune them in memory, and evaluate their own performance. Skills emerge from execution traces rather than human authoring.

MUSE-Autoskill, posted to arXiv on May 27, attacks a problem most production agent systems hand-wave: how do you give an agent a growing library of reusable skills without a human writing every one? The Huawei-led team proposes a closed-loop design with four explicit modules — Skill Creation, Memory, Management and Evaluation — that let an agent extract a new procedure from a successful execution, store it with a description, and decide on later runs whether to recall, refine or retire it.

The Skill Creation step uses an LLM to abstract a successful trace into a parameterized procedure with an entry condition, a body and a postcondition. Memory keeps these in a skill bank with semantic search. Management runs heuristic and learned policies to prune, merge or rerank skills as the bank grows. Evaluation measures whether retrieved skills actually accelerate or improve subsequent runs.

The benchmarks the paper reports show a meaningful sample-efficiency gain on agentic tasks — the same model with MUSE-Autoskill scaffolding improves over 50 tasks versus a baseline that re-derives behaviour every time. The result is most interesting as a paper-and-implementation pair: code is published, and the architecture is small enough to retrofit onto a Claude Code or Codex setup with custom skills.

Why it shows up now

Skill primitives have been the bottleneck for long-horizon agents. Claude Code 2.1.152's reload-skills and disallowed-tools controls landed the same day MUSE-Autoskill posted. The pattern across industry and research is convergent: the next generation of agents are not just better at one-shot tasks, they remember and reuse what worked.

Tech

MobileGym: parallel mobile-GUI-agent simulator for verifiable training

arXiv · 2026-05-27 · Source ↗ ·

An 11-author paper introduces a parallelized mobile-app simulator with deterministic state checks, enabling RL training and evaluation of UI agents at scale.

MobileGym packages a hard, practical problem into a research artifact: training agents that tap, swipe and scroll their way through real mobile apps. The contribution is two-fold. First, the simulator runs mobile-app sessions in parallel with deterministic state inspection, so an RL system can score whether the agent actually achieved the goal rather than relying on screenshot similarity. Second, the team built a verification harness that maps task goals to programmatic post-conditions, which is the part that has tripped up earlier GUI-agent benchmarks.

The paper claims orders-of-magnitude better throughput than serial Appium-style harnesses, with reproducible environments for evaluation. Tasks cover common Android flows — booking a ride, sending money, posting media — across enough app variants to defeat single-app overfitting.

Context

Browser agents have had a Playwright-shaped substrate for a while. Mobile agents have not, which is why production mobile-agent products lag the browser side. MobileGym is the kind of plumbing that lets Apple, Google and the labs evaluate the same kind of agent the OS will eventually ship. The paper landed in the same HF Daily Papers slot as Voxel51's physical-AI survey, which separately argued that the field is underinvested in spatial and embodied benchmarks. MobileGym is one answer in that direction.

Tech

MobileMoE: scaling on-device mixture-of-experts

arXiv · 2026-05-27 · Source ↗ ·

A Meta-affiliated team proposes routing, expert caching, and quantization tricks to run mixture-of-experts models on phones at competitive quality with dense baselines.

MobileMoE goes after the obvious tension in on-device LLMs: dense models large enough to be useful are too heavy for a phone, but smaller dense models give up accuracy. MoE architectures activate only a fraction of parameters per token, which is good for inference cost, but they need fast expert-routing and aggressive caching to be practical on a 12 GB device. The paper proposes a hardware-aware routing scheme that biases toward warm experts, a tiered cache that keeps recent experts in DRAM and the rest in flash, and a quantization recipe tailored to expert sparsity patterns.

The headline experimental result: at the same active-parameter budget as a 3B dense model, MobileMoE matches or beats the dense baseline on standard benchmarks while sustaining usable token-per-second throughput on a phone-class SoC. The paper publishes the routing kernels for ARM-class chips, which is the part that turns the architecture from research into something Pixel and Galaxy teams can copy.

Read with

The on-device story has been gaining momentum since Apple's Foundation Models framework went general availability and Google's Gemini Nano started landing in Android subsystems. MobileMoE makes a credible case that MoE — long thought to be a datacenter-only architecture — can land on phones in 2026, which would compress the latency and privacy advantage of cloud inference.

TechSafety

FinHarness: an inline safety wrapper for finance LLM agents

arXiv · 2026-05-27 · Source ↗ ·

A new paper proposes a lifecycle harness for finance-domain agents that checks pre-trade intent, intra-execution invariants, and post-trade compliance. The system inserts itself between the agent and the broker API.

FinHarness, posted to arXiv on May 27, targets a class of incidents that has been growing alongside the spread of LLM agents into trading and treasury workflows: an agent issues a sequence of orders that look reasonable individually but breach mandate, limit or jurisdictional rules in aggregate. The paper proposes a runtime harness that wraps the agent's tool calls and enforces three checks: pre-action intent verification (does the next action match a declared goal and risk envelope?), in-flight invariant checks (are position limits and counterparty exposures honored after this trade?), and post-action compliance audit (does the resulting state satisfy regulatory reporting and mandate constraints?).

The harness uses a mix of formal specifications and learned classifiers. Formal specs handle hard constraints (no naked short on a banned name, no trade after a kill switch). Classifiers handle softer policy (this looks like wash trading, this looks like a recovery from a mistake versus a real strategy shift). The paper reports a low false-positive rate on synthetic-but-realistic trading streams and documents catch rates on a set of known-bad agent behaviors that earlier monitoring stacks missed.

Why now

The deployment context is unsubtle. JPMorgan, Goldman Sachs and a handful of large hedge funds have moved Claude- and GPT-class agents into research and middle-office workflows in 2026. Anthropic's Project Glasswing security suite was added the same month. The regulators are watching: the SEC's AI-related sweep letters have asked banks to document the supervision regime around any agent that touches the order management system. FinHarness is research, but the architecture maps directly to what compliance teams are now asking vendors to build.

Tech

Guiding LLM post-training with sparse-autoencoder internals

arXiv · 2026-05-27 · Source ↗ ·

A new paper uses features learned by sparse autoencoders on a base model to score and curate post-training data, claiming better downstream task performance than pure heuristic filters.

Post-training data engineering — picking which conversations, completions and synthetic examples to fine-tune on — has been an art with heuristic tooling: length filters, dedup pipelines, classifier-based quality scoring. The May 27 paper takes a different cut. It trains a sparse autoencoder on the base model's internals, gets a set of interpretable features (refusal, hallucination, instruction-following, math-reasoning, etc.), and uses those features to score candidate post-training examples.

The proposed pipeline keeps examples whose feature activations match the desired downstream profile, downweights examples that fire features correlated with bad behavior (toxic, sycophantic, fabricated), and rebalances coverage across feature clusters that the model is weak on. The result is a data-engineering loop that uses the model's own internal representations as the curation signal.

The paper reports gains over standard heuristic filtering on instruction-following and math benchmarks for a frontier-scale base model. It also shows that the sparse-autoencoder features used for scoring are stable across model checkpoints, which makes the technique reusable without redoing the SAE training each release.

Bigger frame

Sparse autoencoders have moved from interpretability curiosity to load-bearing tool over the past eighteen months. Anthropic and DeepMind have both used them for safety steering. This paper is the first prominent attempt to use SAE features for data curation at scale, and it lines up with what frontier labs have been hinting at: the post-training stage now uses the model to evaluate the model.

Tech

KAIST publishes geometry-aware denoising for multi-view 3D reconstruction

arXiv / KAIST AI · 2026-05-27 · Source ↗ ·

A KAIST AI team proposes geometry-conditioned denoising layers that consume camera pose and surface normals during 3D reconstruction from posed images. Tested on standard MV-3D benchmarks.

The KAIST AI group's May 27 paper, which picked up 35 upvotes on HF Daily Papers, attacks robustness in multi-view 3D reconstruction. Standard MV-3D pipelines tend to fail when input images are noisy, motion-blurred, or have aggressive lighting changes. The paper proposes denoising layers that are explicitly geometry-aware — the denoising step consumes camera intrinsics, extrinsics, and predicted surface normals — and a representation that lets the denoiser leverage cross-view consistency to filter true signal from per-image noise.

The team reports gains on ScanNet++ and DTU and shows that the geometry-aware denoising recovers cleanly when adversarial blur is injected into one or two of the input views. The reconstruction quality is competitive with two-stage pipelines but the architecture stays end-to-end.

What it slots into

3D reconstruction sits at the bottom of a long stack — robotics planning, AR rendering, autonomous-vehicle perception. A denoising primitive that holds up against real-world input degradation is one of the missing pieces for letting Cosmos-style world models train on uncurated capture rather than studio data. KAIST has been publishing aggressively in this area; the May 27 paper is the latest in a sequence that includes their physics-aware reconstruction work earlier in 2026.

Tech

EvalVerse: Tencent ships pipeline-aware benchmark for cinematic video generation

arXiv / Tencent · 2026-05-27 · Source ↗ ·

Tencent's video-AI team published a benchmark calibrated against professional editors that scores T2V models on shot composition, narrative continuity and post-production friendliness.

Text-to-video benchmarks have been a mess. They tend to measure superficial fidelity (faces look like faces) and miss the things working filmmakers care about: shot composition, lens choice, continuity across cuts, the friendliness of generated footage to compositing and color pipelines. Tencent's EvalVerse, presented on May 27 with 76 upvotes on HF Daily Papers, builds a benchmark whose scoring rubric was calibrated by a panel of professional cinematic editors.

The benchmark is "pipeline-aware" — generated clips are not just scored as videos but also as inputs to downstream VFX pipelines. Tests include whether the output's motion vectors are coherent enough for stabilization, whether the lighting baked into the clip can be rebalanced, and whether elements can be cleanly separated for compositing. The paper publishes the rubric, the expert calibration data, and a leaderboard comparing major T2V models.

The result reads as a pointed correction to the current top-of-leaderboard claims in T2V. EvalVerse shows that several of the splashier T2V models score poorly when the rubric shifts from "looks impressive in isolation" to "would survive a real post-production session." The release is also a self-interested move: Tencent has Hunyuan-Video competing in this space and benefits from a benchmark that rewards craft over spectacle.

Tech

SpatialBench: is your spatial foundation model an all-round player?

arXiv / Ropedia · 2026-05-27 · Source ↗ ·

A new benchmark from Ropedia evaluates spatial foundation models across reasoning, retrieval and grounded action, arguing today's models specialize too narrowly.

SpatialBench scored 75 HF upvotes on May 27 by making a sharp claim: current spatial foundation models — the kind of model that's supposed to ground language to 3D, plan over a scene, and answer where-questions — overspecialize on one slice of spatial cognition and ignore the rest. The Ropedia team built a battery covering relational reasoning, occlusion handling, navigation, retrieval against scene graphs, and instructable action in simulated environments.

The headline result: most current spatial models do well on one or two slices and badly on the rest. Models tuned for navigation tend to fail at relational counting. Models that excel at scene-graph reasoning tend to break down when motion or occlusion enters the prompt. SpatialBench publishes per-task scores so that builders can pick a model based on the slice they actually need rather than the aggregate.

Adjacent

The release pairs with the Voxel51 physical-AI survey from the same day: the field is rich in claims, thin on benchmarks that survive real-world distributions. SpatialBench, like EvalVerse for video, is the kind of benchmark-as-correction that lands when a category has been overpromising.

Tech

Bearly AI ships OpenADE for structured AI coding

Bearly AI · 2026-05-27 · Source ↗ ·

OpenADE wraps GPT-5.5, Codex and Claude Code in a structured plan-edit-commit loop with automatic git snapshots, run locally with code privacy preserved. The project is open source.

Bearly AI released OpenADE on May 27 as an open-source structured workflow for AI-assisted coding. The pitch is that vibe-coding sessions, while productive, lose state and make it hard to back out of bad changes; OpenADE imposes a planning step, a commented-edit step, and an automatic git-snapshot step that lets the user roll back to any prior state with one command.

OpenADE runs on GPT-5.5, Codex, and Claude Code as model backends. The project reports 82.7% on Terminal-Bench 2.0 with GPT-5.5 — competitive with the headline scores OpenAI quoted at launch — and lower variance run-to-run because the structured planning step makes the model commit to an approach before generating code. The tool runs locally, which addresses the data-leak concern that has kept some teams away from cloud-only coding agents.

Read with

OpenADE is the latest entrant in a small but growing category — Aider, Continue, and now OpenADE — that wraps multiple foundation-model APIs in opinionated structure rather than competing with them. The economics work because the structure improves model performance at the same per-token spend, and the projects can integrate any new model the day it ships.

TechApplication

Celonis launches Context Model, acquires Ikigai for decision intelligence

Celonis · 2026-05-27 · Source ↗ ·

Celonis announced its Context Model — a real-time digital twin of business operations — and acquired Ikigai Labs to roll decision-intelligence capabilities into the same platform.

Celonis announced two moves on May 27. First, a new Context Model that builds a real-time digital twin of an enterprise's operational state — order flows, customer touchpoints, vendor exposures — and exposes it as a structured context for AI applications. The pitch is that LLM-based agents fail at enterprise tasks because they don't actually understand the process they're operating inside; the Context Model is meant to be the missing substrate.

Second, Celonis acquired Ikigai Labs, whose decision-intelligence platform brings forecasting, what-if modeling and constrained optimization into the same toolchain. The combination means that an agent invoked against the Celonis surface can see the present state, simulate a counterfactual, and pick an action with declared cost and risk. Celonis pitches this as the antidote to "blind agents" that issue actions without process awareness.

Frame

The deal lands in a quiet but consequential category — process intelligence — that became central in 2026 as enterprises tried to move AI agents from pilots to production. SAP and Salesforce have been building toward the same destination from the ERP and CRM sides. Celonis is staking the position that the system-of-record-agnostic process layer is where the agent context belongs.

EconomicTech

Tensormesh raises $20M from AMD, CoreWeave to cut inference costs with KV cache reuse

Tech Startups · 2026-05-27 · Source ↗ ·

Seed extension led by AMD Ventures with CoreWeave, NVentures, Valley Capital, and Laude Ventures. The San Francisco startup uses KV caching to claim 10× reductions in latency and GPU spend.

Tensormesh closed a $20M seed extension on May 27 from a notable mix of strategic and venture backers — AMD Ventures, CoreWeave, NVentures (NVIDIA), Valley Capital Partners and Laude Ventures — bringing the company's total raised to $24.5M. The startup is pitching inference infrastructure that exploits KV-cache reuse: in a multi-turn or multi-user system, many prefix tokens are shared across requests, and the attention KV cache for those prefixes can be reused rather than recomputed.

The technical bet is not novel as an idea — vLLM and SGLang have built prefix caching primitives, and llama.cpp has shipped variants since 2024 — but Tensormesh's pitch is that doing it well across heterogeneous GPU pools, with cache eviction policies tuned to enterprise traffic patterns, yields 10× margin in production. The strategic investors say something about where the inference market is going: an AMD-CoreWeave-NVIDIA tag team backing the same inference-optimization startup is rare and signals consensus that the next bottleneck is software-side.

Why now

Coding agents and long-context agentic workflows generate enormous repetition in prefixes. The shift from chat-style assistants to agent-style systems has been the proximate cause for the inference-cost crunch reported by every enterprise vendor over the past quarter. KV-cache reuse turns expensive token recomputation into cheap memory I/O. Tensormesh's claimed numbers are aggressive; the strategic backing suggests at least one large operator believes they hold up.

EconomicTech

RevEng.AI raises $15M led by NATO Innovation Fund

Tech Startups · 2026-05-27 · Source ↗ ·

London-based RevEng.AI raised a $15M Series A with NATO Innovation Fund, Sands Capital, In-Q-Tel, IQ Capital and Episode One. The platform verifies binaries — including those produced by AI coding agents.

RevEng.AI's $15M Series A on May 27 brings the London startup's total to $19.5M and gives it an unusually-aligned investor base: NATO Innovation Fund leading, In-Q-Tel (CIA-affiliated) participating, plus traditional UK and US venture backing. The product does binary-level verification — analyzing compiled software without access to source — to find vulnerabilities, malicious patterns, or backdoors.

The pitch in 2026 is sharper than it was last year. As AI coding agents (Devin, Codex, Claude Code, grok-build) start landing meaningful amounts of code in production binaries, the buyers — defense, intelligence, banks, regulated software vendors — want a way to inspect what actually compiled, not just what the agent claimed it wrote. RevEng.AI's pitch is that source review and unit tests are no longer enough: the AI-shaped supply chain risk lives in the binary.

Adjacent

The investor mix has signal value. NATO-affiliated and In-Q-Tel money does not flow into commodity dev-tools; it flows into things that defense and intelligence buyers expect to procure. RevEng.AI is positioned to be one of the first products lined up against an AI-driven software-supply-chain risk model that policy and procurement bodies are starting to require.

Economic

Fireworks AI in talks for new round at $15B valuation

Bloomberg · 2026-05-27 · Source ↗ ·

Index Ventures, an existing investor, is set to co-lead a new round that would value Fireworks AI at $15 billion. The startup specializes in serving open-source and proprietary models at low cost.

Fireworks AI is in talks for a new fundraising round at a $15 billion valuation, according to Bloomberg's May 27 reporting. Index Ventures, a returning investor from prior rounds, is set to co-lead the new financing. The valuation jump puts Fireworks in the tier of model-serving infrastructure businesses that became darlings of the inference economy after Cerebras, Groq and Together AI scaled their early-2026 rounds.

Fireworks differentiates by tuning open-source models (Llama, Qwen, DeepSeek, Mistral OSS) for production traffic and serving them at lower cost than the labs charge for their proprietary models. Their customer base includes companies running coding agents, retrieval-heavy chat products, and other token-intensive workloads where moving even a fraction of traffic off frontier APIs saves real money.

Frame

The Fireworks talks and the Tensormesh seed extension on the same day are different bets on the same insight: inference is now the primary cost line for AI applications, and the labs are still charging the most. Fireworks attacks the problem at the serving layer (different model, same task, lower cost). Tensormesh attacks it at the infrastructure layer (same model, smarter caching, lower cost). Both are seeing investor demand because the cost curve matters more than the capability frontier for buyers who already have something that works.

Economic

Slamcore raises $14M for industrial spatial intelligence

Tech Startups · 2026-05-27 · Source ↗ ·

London's Slamcore closed a $14M round led by ROKStar Ventures, the venture arm of Rockwell Automation. The stereo-camera SLAM platform tracks industrial vehicles without GPS or facility modification.

Slamcore announced a $14M round on May 27, bringing total funding to $40M. ROKStar Ventures — the venture arm of Rockwell Automation — led, with Toyota Ventures, Interwoven Ventures, MMC Ventures, Amadeus Capital Partners and IP Group participating. The company sells a stereo-camera-plus-visual-AI system that tracks forklifts, AGVs and other industrial vehicles in warehouses and factories without requiring GPS, fiducials, or any modification to the facility.

The industrial buyer logic is straightforward. Logistics and manufacturing operations need to know where their fleet is at sub-meter accuracy in indoor environments where GPS fails. Existing solutions either require expensive infrastructure retrofits (UWB beacons, magnetic strips, RFID portals) or operate at coarser accuracy than the workflow needs. Slamcore's pitch is that better visual SLAM, running on a cheap stereo camera, hits the accuracy bar without the retrofit cost.

Adjacent

Voxel51's State of Visual and Physical AI Survey on the same day put numbers on the broader market: 78% of practitioners see value, 74% say the field is underinvested, and 92% identify world models and spatial intelligence as the next frontier. The Rockwell-led investment in Slamcore reads as one of the strategic responses — industrial automation buyers betting that the spatial-AI moment is arriving even if VC dollars are still flowing harder into LLM startups.

Economic

ClearNote Health raises $52M for AI-plus-epigenomics cancer detection

Tech Startups · 2026-05-27 · Source ↗ ·

San Diego-based ClearNote Health raised $52M led by Mattias Westman, taking total funding past $185M. The company combines methylation epigenomics with AI for early cancer detection.

ClearNote Health, the San Diego oncology-diagnostics company spun out of Stanford and Salk-affiliated labs, raised $52M on May 27 led by Mattias Westman and an unnamed global long-only manager. The round brings ClearNote's total funding past $185M. The platform combines blood-based methylation epigenomics with AI-driven signal extraction to detect cancer earlier than standard tumor-marker assays.

The company has been moving along the regulated-diagnostics path rather than chasing volume in the wellness market. ClearNote's pancreatic-cancer assay is in clinical use, and the new capital supports continued data generation and the expansion into a multi-cancer detection platform. The AI piece is real: extracting weak cancer signal from methylation patterns at parts-per-million sensitivity requires models that can integrate noisy genomic, epigenomic and clinical context.

Frame

Funding for AI-in-diagnostics has been quieter than funding for AI-in-coding, but the strategic investors are different — public-equity-style long-onlys, biotech operators, and tier-1 health systems. The model is closer to selling lab tests than to selling SaaS. ClearNote's round suggests the category is still drawing meaningful capital despite the headline attention going to model companies.

ApplicationPolicy

DataGrail report: data processing agreements no longer trustworthy at face value

DataGrail · 2026-05-27 · Source ↗ ·

DataGrail's annual Privacy and AI Trends Report finds that vendor DPAs increasingly fail to describe what AI features actually do with the data they touch, and many enterprises are missing the gap.

DataGrail's 2026 Privacy and AI Trends Report, released on May 27, came to a sharp conclusion: the data processing agreements (DPAs) that enterprises rely on to bound vendor data use have not kept up with what vendor AI features actually do. The report finds that AI-augmented features inside enterprise SaaS often process customer data in ways — training auxiliary models, building tenant-level retrieval indices, sharing across product surfaces — that the DPAs do not describe in operative detail.

The practical consequence: a buyer relying on the DPA as the safety boundary may be exposed to data uses they did not understand and did not authorize. The report calls for buyers to require active, technical, vendor-by-vendor verification of AI feature behavior rather than treating signed paper as proof. DataGrail surveyed privacy and AI-governance leaders and found that 81% of respondents had encountered at least one vendor AI feature that surprised them after enablement.

Read with

The findings line up with what regulators are now publicly saying. Canada's federal Privacy Commissioner ruled in early May that OpenAI's training-data collection violated PIPEDA and provincial laws. Several U.S. state attorneys general are pursuing similar lines. The DPA-trust gap that DataGrail documents is the operational symptom of the same regulatory shift.

Application

VentureBeat: AI coding benchmarks have misled enterprise buyers

VentureBeat · 2026-05-27 · Source ↗ ·

A May 27 piece argues that the headline coding benchmarks used in enterprise procurement (SWE-Bench, Terminal-Bench) cluster the frontier models too tightly and miss the practical differences buyers actually feel.

Michael Nuñez's piece in VentureBeat on May 27 makes a careful argument that enterprise buyers picking coding models on SWE-Bench Verified and Terminal-Bench scores have been steered toward an over-tight comparison. GPT-5.5, Claude Opus 4.7, and Gemini Pro all land within a handful of percentage points on those benchmarks, but the developer experience and the cost-per-task they produce in actual repositories vary dramatically.

The article catalogs three failure modes of benchmark-driven procurement: (1) benchmarks measure isolated task completion but not multi-step recovery from failure, which is where real codebases live; (2) cost-per-completed-task can vary 5× across models that score within 2% on benchmarks, because token efficiency and tool-use patterns differ; (3) the benchmarks are increasingly part of the training distribution, eroding their predictive power for new buyers.

Why now

The piece lands the same day Cognition raised $1B at $26B, and the same week Anthropic, OpenAI and xAI shipped competing coding-tool updates. Enterprise buyers in 2026 are evaluating coding-agent vendors with budget allocations measured in tens of millions. The argument that the published numbers are a misleading scorecard is timely and uncomfortable.

Policy

Microsoft Marketplace Terms updated, effective May 27 with no opt-out

Microsoft Partner Center · 2026-05-27 · Source ↗ ·

A change to the Microsoft Marketplace Terms of Use took effect automatically on May 27 with no further action required by any party. The update covers AI-services marketplace transactions.

Microsoft updated its Marketplace Terms of Use with an automatic effective date of May 27, 2026. The update is the latest in a sequence of Microsoft policy moves around the Azure Marketplace, AppSource, and the channels through which AI services and agents are now distributed at enterprise scale. The change applies without partner action and affects every active publisher and reseller agreement.

The update is operational rather than dramatic — it tightens language around evaluation transactions, refines the definition of "Microsoft Pay" for marketplace billing, and clarifies how AI-services revenue is attributed in commission and partner-incentive calculations. The lack of opt-out is the consequential part: marketplace participants do not get to negotiate or refuse.

Frame

Microsoft's marketplace is one of the primary distribution surfaces for enterprise AI vendors — Anthropic models on Azure, third-party agents, OpenAI tools, custom Copilot extensions. The terms governing that surface have moved several times in the past twelve months as Microsoft has navigated its restructured OpenAI partnership and the addition of Claude and other competing models to Azure. May 27 is one more turn of the same wheel.

Safety

Claude Code patches PowerShell permission-bypass vulnerability

Anthropic · 2026-05-27 · Source ↗ ·

The Claude Code 2.1.152 release notes flag a critical PowerShell permission-bypass fix. Earlier builds let certain PowerShell invocations execute outside the consented permission scope.

Buried in the Claude Code 2.1.152 release notes is a fix Anthropic flagged as critical: a PowerShell permission-bypass vulnerability in earlier builds let certain PowerShell invocations run outside the permission scope the user had consented to. In a CLI that already has permission to read and write files in the working directory, this is the difference between a tightly-bounded session and a session that could touch the rest of the user's machine.

Anthropic did not publish a CVE number with the release notes, and the description does not detail the exploitation path. What is clear from the notes is that the patch is in 2.1.152 and that earlier builds should be upgraded. The release also includes a fix for terminal styling degradation in long sessions ("recycling the renderer's style pool") and several quality-of-life improvements.

Wider context

Coding-agent CLIs are increasingly running with broad shell privileges by default. The combination of "Auto mode no longer requires consent" — also in 2.1.152 — and shell-execution permissions makes the integrity of the permission boundary critical. The PowerShell fix is timely. The broader question for the category is whether agent CLIs are converging on a permission model strong enough to bound the consequences of any single vulnerability.

Trend

Simon Willison: Anthropic and OpenAI have hit product-market fit

Simon Willison · 2026-05-27 · Source ↗ ·

Willison argues that both labs crossed into genuine product-market fit through coding agents, with enterprise pricing moves in April and rumored profitability at Anthropic as the proof.

Simon Willison's blog post on May 27 makes the case that Anthropic and OpenAI have moved past the consumer-popularity moment and into actual product-market fit, with coding agents as the demand driver. The post catalogs the inflection points: November 2025, when Claude Code, Cursor, Codex and Devin all became genuinely useful at multi-step engineering work; April 2026, when both labs restructured enterprise pricing from flat-rate seats to API-metered token consumption, conceding that the seat model would be uneconomic for the new workloads.

The financial evidence Willison pulls together is concrete. Anthropic is rumored to have reached $10.9B quarterly revenue and first profitability. SpaceX's S-1 filing, which leaked earlier in May, disclosed a $1.25B per month Anthropic compute commitment running through May 2029. OpenAI now has 703 open positions, 32.6% of which are enterprise-focused; Anthropic has 390 open positions, 26.9% enterprise-focused. Willison's personal usage — $1,199 to Anthropic and $980 to OpenAI in a 30-day window under nominally $200/month subscriptions — illustrates the unit-economics shift at the customer level.

"You'd need 1-2 billion subscribers sticking around for four years to cover [$1 trillion in infrastructure]."— Simon Willison, on why coding agents matter more than ChatGPT subscriptions for paying off AI infrastructure

Why this matters today

The post lands as Cognition raised $1B at $26B on the same day, with $492M run-rate. If Willison's read is right — that the AI business is now real because coding agents are real — then Cognition's funding is a venture-market validation of the same thesis at the pure-play level, and the price-war moves from Anthropic, OpenAI, Google and xAI in May are the labs reacting to the same dynamic.