Archive
27 published pieces, newest first.
2026
Compute & power · The Stack · June 28, 2026
SGLang ships day-zero MoE kernels faster than most teams pick a runtime
Apache-2.0 serving framework. v0.5.14 landed 26 June 2026 with DeepSeek-V4 and Waterfill/LPLB load balancing.
Policy & law · Dispatch · June 17, 2026
European Parliament votes to postpone high-risk AI Act duties
16 June 2026 plenary: 423-57-174. Stand-alone high-risk systems move to 2 December 2027; embedded safety-component systems to 2 August 2028. Official Journal in-force date is UNKNOWN.
Agents & software · Dispatch · June 9, 2026
Apple opens Siri AI to developers and withholds it from EU iPhone
8 June 2026 WWDC: developer testing now, user beta later this year. DMA update: no Siri AI on iOS 27, iPadOS 27, or paired watchOS in the EU. macOS 27 and visionOS 27 still get it.
Policy & law · Dispatch · June 3, 2026
EO 14409 starts 30- and 60-day clocks on covered frontier models
2 June 2026 order: CISA guidance, a Treasury vulnerability clearinghouse, and a voluntary pre-release access path. Section 3(c) says it is not a licensing regime.
Deployment · Field Notes · June 2, 2026
EU AI Act transparency war room runs while high-risk clock moves to 2027
Digital Omnibus (EU) 2026/1744 in force 27 July; Article 50 transparency from 2 August; Annex III high-risk deferred to 2 December 2027.
Compute & power · Dispatch · June 1, 2026
NVIDIA says Vera Rubin is in full production
10x vs Grace Blackwell, SPDX of the MGX design, export licenses, and a firm shipment week are UNKNOWN.
Models · Dispatch · May 29, 2026
Anthropic ships Claude Opus 4.8 at $5/$25 with cheaper fast mode
28 May 2026. Same list price as Opus 4.7. Fast mode $10/$50. System card: not past Mythos Preview on RSP. Prompt-injection tests vs 4.7 are mixed.
Agents & software · Ground Truth · May 25, 2026
JSON Schema makes agent tools production-ready
NIST AI 800-5 finds commenters agree agent threats are novel and existing controls need adaptation. Schema syntax is not the bar.
Models · The Record · May 22, 2026
NIST AI 800-5 synthesis finds agent threats novel but cybersecurity principles adaptable
Published May 18, 2026; summarizes CAISI RFI NIST-2025-0035 (comments due March 9, 2026).
Policy & law · Dispatch · May 16, 2026
Colorado enacts SB26-189, replacing its AI Act with ADMT notice rules
Official bill page lists the measure as enacted. Duties start 1 January 2027 if the attorney general finishes rules. June 2026 high-risk AI deadlines from SB24-205 are not the live calendar.
Models · Ground Truth · May 8, 2026
One million tokens fixes long-document recall
DeepSeek-V4 reports MRCR 1M MMR 83.5 and CorpusQA 1M ACC 62.0 at Pro Max. Strong on paper. Not a buyer corpus guarantee.
Compute & power · The Record · May 5, 2026
DeepSeek-V4 paper reports 27% FLOPs and 10% KV cache at 1M tokens versus V3.2
arXiv:2606.19348 (April 26, 2026); MIT checkpoints on HuggingFace; CSA/HCA hybrid attention.
Work & money · Field Notes · April 30, 2026
McKinsey skill-partnership math reframes headcount fights as workflow redesign
MGI report: ~57% of US work hours theoretically automatable today — a capability ceiling, not a job-loss forecast; April 2026 podcast walks the partnership frame.
Policy & law · Dispatch · April 28, 2026
DeepSeek-V4-Pro ships on Hugging Face under MIT, not a community carve-out
arXiv 2606.19348 and the model card both cite MIT on weights and repository code. Export and deployment diligence still sit with the buyer.
Policy & law · Dispatch · April 10, 2026
BIS pushes Approved IC Designer applications to December 31, 2026
Federal Register 2026-06851 replaces the April 13, 2026 cutoff in Note 1 to ECCN 3A090.a. The rule took effect April 7.
Work & money · Dispatch · April 2, 2026
OpenAI closes $122 billion round at an $852 billion post-money valuation
31 March 2026 company post: committed capital, not a cash-in-bank figure. Amazon, NVIDIA, and SoftBank anchored after a 27 February $110 billion step.
Work & money · Ground Truth · March 28, 2026
Enterprise AI is paying off at scale
McKinsey's 2026 AI Trust survey finds average RAI maturity 2.3 and only ~30% at level 3+ on governance. Not an EBIT proof point.
Models · The Stack · March 28, 2026
SAM 3.1 multiplexes video objects; license is Meta SAM, not Apache
27 March 2026 checkpoints on Hugging Face `facebook/sam3.1`. Meta reports ~7x at 128 objects on one H100 versus SAM 3. Weights gated. GitHub stars UNKNOWN.
Agents & software · Field Notes · March 25, 2026
Vertex Gen AI eval pipeline replaces RAG demo scorecards
Google’s Vertex evaluation service and EvalTask score retrieval and generation separately; model-based rubrics replace hallway comparisons.
Deployment · Field Notes · March 12, 2026
Copilot agents split on declarative vs custom engine before Control System gates
Microsoft docs: declarative agents inherit M365 compliance; custom engine agents bring your own orchestrator and hosting bill.
Models · Dispatch · February 20, 2026
Google ships Gemini 3.1 Pro in preview across app, Vertex, and Copilot
19 February 2026. Google cites a 77.1 percent ARC-AGI-2 score. General availability date is UNKNOWN. GitHub Copilot Business and Enterprise admins must flip a policy.
Models · Dispatch · February 18, 2026
NIST CAISI opens an AI Agent Standards Initiative on three pillars
The 17 February 2026 launch ties industry-led protocols to agent security research. An RFI on agent security closed 9 March.
Compute & power · Dispatch · January 16, 2026
H200 and MI325X China licenses go case-by-case, still not a general license
BIS final rule 2026-00789, effective 15 January 2026, reviews U.S. exports of those SKUs case by case if exporters certify U.S. supply, a 50 percent TPP cap, KYC, and U.S. third-party testing. Reexports stay denied.
Models · The Stack · January 15, 2026
MedGemma 1.5 4B adds CT, MRI, and slide patches under HAI-DEF terms
Google Research, 13 January 2026. Weights are not Apache-2.0. Clinical diagnosis is out of intended use. Internal accuracy deltas are Google's.
Security · Dispatch · January 14, 2026
NIST CAISI agent-security RFI closed March 9; summary analysis landed in May
Docket NIST-2025-0035 opened 12 January 2026. NIST Trustworthy and Responsible AI 800-5 summarizes responses published 18 May 2026.
Policy & law · Dispatch · January 3, 2026
Texas TRAIGA took effect 1 January 2026; the Attorney General may investigate and sue
House Bill 149, 89th Regular Session. Prohibited uses, government and health disclosures, a sandbox, and a state AI council. Attorney General website mechanism due by 1 September 2026.
Policy & law · Dispatch · January 3, 2026
California SB 53 is in force: frontier labs must publish an AI framework
Chapter 138, effective 1 January 2026. Large frontier developers owe a public frontier AI framework and a transparency report at model deploy. The statutory compute cutoff is not restated from secondary trackers.