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
The 31 May GTC Taipei post puts shipments in “this fall,” names Dell, HPE, Lenovo, and Supermicro, and still says products are when-and-if-available.
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.