2026 - Page 46 of 1450

california data center laws energy water rules a california data center insulator stack ribbed discs

California Tightens Rules on AI Data Center Energy and Water Use

Governor Newsom signed seven data centre bills on 21 September 2026, covering electricity reporting, water disclosure under penalty of perjury, cost causation for grid and water upgrades, and the removal of CEQA categorical exemptions. Read as enacted, the package carries a 10 megawatt reporting floor, a 25 megawatt cap on the tariff threshold, binding deadlines in 2028 and a first published load assessment in 2029. SB 887 also creates a fast track with eleven conditions, including hourly zero-carbon matching within five years. This article goes through all seven statutes and the numbers inside them.

Read more
ai bookkeeping tabby accountants obsolete a ai bookkeeping pot of notched tally sticks

With Tabby, a Former Accountant Is Using AI to Make Accountants Obsolete

TechCrunch profiled Tabby on 21 September 2026 as the AI bookkeeping startup built by a New York CPA to automate his own profession away. The company reports 5,500 businesses, roughly $100,000 in annual recurring revenue and a team of seven. Its own website advertises a separate product for accounting firms at $15 per client per month, a free firm licence, a referral commission and 100+ firms already signed up. This article divides the reported numbers, reads the vendor pages against the headline, and sets out what bookkeeping automation genuinely replaces.

Read more
math advisory group openai 100 open problems a math advisory group amphitheatre concentric steps

OpenAI Forms Math Advisory Group as Its AI Resolves More Than 100 Open Problems

OpenAI announced on 21 September 2026 that an internal model has resolved more than 100 long-standing open mathematical problems, and that an independent advisory group hosted at the Institute for Advanced Study will now advise on how such results are assessed and released. Ten days earlier, Fields Medallists published a declaration calling the goals of AI companies and mathematicians “severely misaligned.” This article checks both name lists, sets out the compute figures OpenAI published for the Navier-Stokes proof, and identifies the one thing the advisory group is explicitly forbidden from advising on.

Read more
anthropic playground workbench claude api testing a jointer plane lying flat with a closed loop handle

Anthropic Rebrands Workbench as Playground for Claude API Testing

Anthropic retired the legacy Workbench on 17 August 2026 and shipped Playground the next day. The replacement is stateless: no saved prompts, no version history, no variables, no evals and no sharing, with nothing stored on Anthropic’s servers. In exchange it supports every Messages API parameter, ships templates for server-side features such as code execution and web search, shows the full SDK request alongside the response for every run, and exports to code. Three experimental prompt endpoints were retired alongside it, and the export window for Workbench data closed permanently on 1 September 2026.

Read more
agentic shopping amazon bars meta muse ai a handcart with two large side wheels and an upright handle frame

Amazon Bars Meta’s Muse AI From Shopping on Its Site

Amazon has cut Meta’s twelve-day-old Muse agent off from shopping on Amazon.com, serving users a popup that cites its Conditions of Use rather than accusing anyone of hacking. That wording is deliberate: on 4 August 2026 the Ninth Circuit ruled in the Perplexity Comet case that the user, not the AI company, is the party accessing a retailer’s computers, and denied rehearing on 10 September. This article covers Amazon’s stated objections, how Muse actually reaches a site with no agent API, the litigation timeline, the $68 billion advertising business at stake, and how Amazon’s own Buy for Me feature differs.

Read more
apple home hub siri ai homepad almost ready a steamer of three stacked round tiers with a domed lid

Apple Home Hub Device Built Around Siri AI May Be Almost Ready

Bloomberg’s Mark Gurman reports that Apple’s long-delayed smart display, codenamed J490, is now in widespread testing in employees’ homes and could be introduced as early as October 2026. It carries a squarish seven-inch screen on a half-HomePod-mini base or a magnetic wall mount, a camera that recognises who is looking at it, and a new operating system built on tvOS foundations and organised around Siri AI. Two refreshed companions, an Apple TV and a HomePod mini, are reported to arrive with it, and a 2027 in-home security system tied to a monitoring service sits behind them. This article separates what is reported from what Apple has confirmed.

Read more
ai gender bias womens language workplace prompts a gramophone with a flared horn on a square box body

AI Might Be Making Women Sound Bad at Work

Johns Hopkins researchers took 427 real workplace writing prompts, rewrote each into a women-associated and a men-associated version, and ran both through GPT-4, Llama, Mistral and Gemma. Every model returned shorter, plainer, less formal documents for the women-coded phrasing. Adding a male or female sign-off name changed nothing. Linear probes decode the linguistic register at 0.988 accuracy by layer 5 while name gender reaches only 0.717, and activation patching puts the causal weight in layers 0 to 7. This article covers the method, the per-model results, the two controls that rule out mirroring, and why the authors say users cannot fix it themselves.

Read more
ai agent manipulation stubborn agents multi agent networks a cartwheel lying flat with a thick hub and eight spokes

When Stubbornness Becomes a Security Problem: How AI Agents Can Manipulate Others

A CISPA Helmholtz Center study borrowed the Friedkin-Johnsen opinion model from sociology, pointed it at six families of large language model, and found that a single stubborn, persuasive agent can hijack an entire multi-agent network. With the attacker at the hub of a star topology, attack success rate reached 1.00 on Gemini-3-Flash and 0.99 on GPT-OSS-120B, against unattacked baselines of 0.02 and 0.06. This article covers the mechanism, the per-model and per-topology numbers, the three structural defences the maths predicts, the trust-adaptive mechanism the team built, and the adaptive attacker that pushes warm-up trust to 0.95.

Read more
mac studio m5 ultra review ai graphics premium a mac studio m5 anvil with a tapering horn

Apple Mac Studio (M5 Ultra) Review: Huge AI and Graphics Power at a Huge Premium

The first review of a shipping Mac Studio with M5 Ultra scored it 8.9 out of 10 and told most buyers to purchase the cheaper model. This analysis separates Apple’s “up to” multipliers — 4.3x peak AI compute against M3 Ultra, but only 1.3x multithreaded CPU — from the measured results: 52,185 on Geekbench 7 multi-core, a 33% gain over M3 Ultra, and only about 10% over the M5 Max on synthetic AI tests. It works through the full configuration ladder from $2,499 to $11,299, the 37.5% entry-price increase, the untested four-machine clustering claim, the whole-system cost per gigabyte of unified memory against an NVIDIA DGX Spark, and the two conditions under which the Ultra is the rational purchase.

Read more
siri ai settlement claims apple 250 million a siri ai settlement piggy bank with a back slot

iPhone Owners Can Now Submit Claims in Apple’s $250 Million Siri AI Settlement

Apple’s $250 million settlement over the Siri features it demonstrated in June 2024 and did not ship until September 2026 is now accepting claims. Roughly 37 million US iPhone owners are eligible on seven models bought between 10 June 2024 and 29 March 2025, with a presumptive payment of $25 per device and a $95 ceiling. This guide covers the eligibility rules, the serial number requirement, the arithmetic that decides where in that range the payment lands, the court timetable through the 24 February 2027 final approval hearing, what Apple did and did not concede, and the questions the settlement leaves open — including everyone outside the United States.

Read more
CHAT