The AI Dispatch — May 4, 2026

Wall Street

Anthropic, Goldman, and Blackstone Launch $1.5B AI Services Joint Venture

A formalized race to own the services layer between frontier labs and the Fortune 5000: portfolio-company AI engineering, built around Claude agents, with a near-identical OpenAI counterpart announced the same day.

By AI Dispatch Desk · May 4, 2026 · Source: CNBC

Anthropic, Goldman Sachs, Blackstone, and Hellman & Friedman jointly announced on Monday morning a $1.5 billion AI services firm structured around embedding engineering teams inside mid-sized portfolio companies and redesigning their core operating workflows around Claude agents. The venture, which carries no consumer-facing brand at launch, is positioned as the connective tissue between Anthropic’s frontier models and the thousands of investor-owned mid-market firms that lack the in-house technical depth to deploy agentic systems on their own.

The founding partner roster goes well beyond Goldman and Blackstone. General Atlantic, Apollo Global Management, Sequoia Capital, Leonard Green & Partners, and Singapore’s sovereign wealth fund GIC all signed on at announcement — a coalition that, taken together, represents portfolio exposure across hundreds of companies. The structural implication is straightforward: the venture is being launched with a built-in client pipeline at scale, sidestepping the slow customer-acquisition cycles that have characterized enterprise AI services to date.

Within hours of the Anthropic announcement, a near-identical competing structure was reported at OpenAI, backed by TPG and Bain Capital, with comparable founding capital and a comparable mandate. The simultaneous emergence of two parallel ventures — one Anthropic-aligned, one OpenAI-aligned, both anchored by tier-one private equity sponsors — formalizes what had until now been an informal scramble. The AI-services layer between frontier labs and enterprise customers is no longer being built by consultancies repurposing their existing practices. It is being built ground-up as a distinct competitive arena.

The economic logic for the model labs is direct: services revenue is sticky, defensible, and creates contractual commitments to specific model families that pure API consumption does not. For the private equity sponsors, the bet is that portfolio-company productivity gains from properly deployed AI agents will compound into materially better exit multiples within the typical fund holding period. The losers, if the model works as designed, are the traditional consulting firms whose AI services revenue depended on positioning between the labs and the customers — a position both new ventures are constructed to eliminate. Press materials indicate the firm will be operational by the end of the quarter, with initial engagements already under contract.

The Monday Bundle

Research Calendar, State Laws, and a Reshuffled Leaderboard

NeurIPS deadlines drive an arXiv surge; Connecticut’s AIRT Act lands on the governor’s desk; and SWE-bench Verified widens the frontier-versus-open gap.

Research Calendar

NeurIPS 2026 Abstract Deadline Triggers arXiv Surge

Source: NeurIPS 2026 Call for Papers

The NeurIPS 2026 abstract registration deadline (May 4, AOE) and the corresponding full-paper deadline two days later (May 6, AOE) create the period’s single largest research news cluster, with thousands of submissions expected across the main track, the datasets and benchmarks track, and the position-paper track. NeurIPS partnered with Google this cycle on an optional Paper Assistant Tool (PAT) that gives authors automated private feedback on submissions before the final deadline — a first for a top-tier machine learning conference and an early signal that LLM-assisted scientific writing is becoming part of the formal publication infrastructure. The conference itself is scheduled for December 2026, but the early-May submission window is what shapes the late-spring arXiv firehose: expect a measurable jump in daily preprint counts through midweek as withdrawn-and-resubmitted drafts hit the public record.

State Laws

State AI Legislative Tracker: CT, MD, Multi-State Chatbot Bills

Source: Troutman Pepper Privacy

Troutman Pepper’s May 4 state legislative update flags Connecticut’s AIRT Act — a comprehensive private-sector AI governance bill — now on Governor Ned Lamont’s desk for signature. In Maryland, Governor Wes Moore signed the Protection From Predatory Pricing Act (HB 895) into law, extending consumer-protection guardrails around algorithmic pricing systems. Chatbot disclosure bills advance in multiple states. The tracker explicitly excludes government-use, insurance, and election-related bills — meaning the private-sector AI governance obligation surface area is broader than the headline counts suggest. The throughline of the report: state-level AI compliance burden is expanding at a pace federal action shows no sign of matching, and the patchwork is starting to look durable.

Leaderboards

SWE-bench Verified Reshuffles as Frontier Pulls Further From Open

Source: BenchLM

The SWE-bench Verified leaderboard saw notable movement in early May as NeurIPS submissions coincided with new system drops from the frontier labs. Claude Mythos Preview — available only under restricted access, and powering Project Glasswing — holds the top slot at 93.9%. GPT-5.5 is at 88.7%. Claude Opus 4.7 Adaptive comes in at 87.6%. The frontier-versus-open gap has widened, not narrowed: no fully open-weight system has crossed 70% on the verified 500-problem subset this cycle, despite a wave of new open releases timed around the conference deadline. The implication for the open ecosystem is a familiar but sharpening one: closed-frontier systems are extending their lead on tasks that matter most for enterprise procurement decisions, even as open systems continue to close gaps on academic benchmarks.

Conference Roundup

ICML Camera-Ready Pulls Accepted Papers Into Public View

Source: ICML 2026 Dates

With ICML 2026 camera-ready submissions due May 28 and the main conference scheduled for July 6–11 in Seoul, a parallel wave of accepted papers is appearing as updated arXiv preprints in the same window as the NeurIPS abstract surge. ICML 2026 has also imposed its strictest-ever authorship policies, including an explicit ban on listing large language models as authors. The combined effect of the two conference calendars is that early May functions as the year’s densest single window for serious machine-learning research, with most labs timing their highest-confidence releases to land while the academic community is at peak attention.

Toolbox

Claude Code’s May 4–8 Sprint: Plugin Guards, Windows PowerShell, Marketplace Token Estimates

By AI Dispatch Desk · May 4, 2026 · Source: Release Bot — Anthropic Claude Code

Claude Code versions v2.1.128 through v2.1.136 are scheduled to land across the week, with a recurring theme: harden the plugin system, smooth the Windows developer experience, and surface cost information earlier in the marketplace flow. The headline changes the team is shipping:

Plugin and Marketplace

Plugin dependency enforcement: claude plugin disable now refuses to disable a plugin when another installed plugin depends on the target, and enable auto-resolves transitive dependencies before activation
Projected per-turn token cost estimates added to the /plugin marketplace browse pane, so users can see expected spend before installing a plugin family
Marketplace browse pane now surfaces last-update timestamps and explicit author-attribution metadata on the card view

Windows, MCP, and Reliability

PowerShell defaults to -ExecutionPolicy Bypass on Windows and is enabled by default — eliminating a long-standing onboarding friction for Windows users
Fixed MCP_TOOL_TIMEOUT not raising the per-request fetch timeout for remote HTTP and SSE MCP servers (the variable previously only affected stdio transports)
Background session reliability: fixed a daemon disconnect after macOS sleep/wake cycles that had been silently terminating long-running agent runs
Improved error messaging when a plugin manifest declares a dependency that cannot be resolved from any configured marketplace

Taken together, the sprint reads as the inflection point at which Claude Code’s plugin system stops being a power-user feature and starts being treated as a load-bearing primitive: dependency enforcement, cost transparency, and platform-parity polish are the kinds of changes that ship when a feature is graduating from beta-grade to assumed-stable.

Briefs

Across the Wire

Two short items from the research and conference circuit — a convergence theorem for sparse autoencoders, and an ICML camera-ready wave with the strictest LLM-authorship policy yet.

SAE Theory Convergence: Piecewise Biconvexity and Spurious Minima

“A Unified Theory of Sparse Dictionary Learning in Mechanistic Interpretability: Piecewise Biconvexity and Spurious Minima” (arXiv 2512.05534) received a substantial revision on May 2, formalizing the convergence conditions under which sparse autoencoders — the workhorse instrument of modern mechanistic interpretability — reliably recover the underlying feature dictionary. The revision sharpens earlier statements about when SAE training is well-posed and when it is not, with practical implications for any interpretability program that depends on SAE-derived feature explanations being stable across runs. The result is one of a small but growing set of papers giving the SAE literature its theoretical foundations after a long stretch of empirically driven work.

ICML 2026 Camera-Ready Wave Lands With Strictest LLM-Authorship Policy Yet

With ICML 2026 camera-ready submissions due May 28 and the conference itself set for July 6–11 in Seoul, a wave of accepted papers is appearing as updated arXiv preprints across the first half of May. The conference has imposed its strictest-ever submission and authorship policies this year, including an explicit ban on listing LLMs as authors and tightened disclosure requirements for AI-assisted writing. The policy combination is being read as the first serious institutional response from a top-tier venue to the post-LLM authorship-and-attribution problem — an issue that NeurIPS’s simultaneous opt-in PAT pilot is approaching from the opposite direction.

GitHub Trending

Monday Snapshot — the repositories moving the needle on the first trading day of May.

GitHub Trending — Monday Snapshot
Repo	Language	Stars	What it does
mattpocock/skills	TypeScript	+44.5K May	Curated collection of reusable Claude Code skills; the de facto registry for the agent-skills ecosystem.
multica-ai/andrej-karpathy-skills	Markdown	+35K May	Karpathy-attributed Claude Code skill set distilled from his teaching corpus into actionable agent skills.
nexu-io/open-design	TypeScript	+38K May	Local-first design-system generator targeting Claude, Codex, and Cursor as backends — no cloud dependency required.
antoinezambelli/forge	Python	~91K	Reliability framework for self-hosted LLM tool-calling: retry policies, schema validation, and structured-output recovery.
astral-sh/uv	Rust	~85K	Fast Python package manager and project tool, now standard in the ML toolchain ahead of the NeurIPS submission rush.
colbymchenry/codegraph	TypeScript	New	Pre-indexed code knowledge graph for Claude Code: lets agents traverse symbol-level relationships without re-reading source.