Andy Wood (@andyhq) header
Andy Wood (@andyhq) profile picture

AI that works. News that's ranked. No filler. https://teapaw.org/t.co/VrgZrPJPUZ

UK
Joined July 2009
Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)Photo by Andy Wood (@andyhq)
Nearly every UK data centre operator says electrical disruption is now a critical bottleneck, according to a new report. Anyone planning capacity, or relying on hosted infrastructure, should be asking how their power is secured. 2 outlets are on it. alluktech.com/infrastructure… #uktech
4
TfL has taken the Oyster photocard website offline after attackers got into customer accounts. Around 265,000 customers may be affected, so check your account and watch for odd activity. 2 outlets are on it. #uktech alluktech.com/security.html?…
13
Questions to ask a supplier who says their AI product is "compliant". 1. Compliant with what, specifically? Ask for the law or standard by name. 2. In which countries? 3. Who checked, and can we see the result? 4. What does that leave for us to do as the firm using it? A supplier's compliance claim may not cover how you use the tool. Telling customers, human checks, record keeping and staff training may still sit with you. A good supplier will have a clear answer. A vague one tells you something too.
24
UK Regulators Propose New Cyber Resilience Guidance as Attacks Surge Expect more cyber resilience checks from UK regulators, so founders and developers should start documenting their security controls and incident response plans now. 8 outlets are covering this. alluktech.com/policy.html?ut…
15
The Digital Hum...
9
Frankenstein's monster got a lot less venture capital and nobody asked it to write a LinkedIn post... UK MPs call for a watchdog that can stop dangerous AI models In its report "Guardrails for Growth: Making AI Safe", a Commons committee chaired by Liam Byrne calls for an AI Safety and Security Bill that would give a regulator power to halt high-risk training, block deployments and require incident reporting. Byrne said AI leaders can't have "unfettered freedom to behave like modern-day Dr Frankensteins", but ministers haven't committed to legislate. theregister.com/ai-and-ml/20…
37
Four levels of permission an AI tool can have. Worth knowing which one each tool in your firm is at. 1. Suggests. You read it and decide. 2. Drafts. It writes, you edit and send. 3. Acts with approval. It prepares the action, you click to confirm. 4. Acts alone. It sends, pays, books or deletes without a click. This describes what a tool is allowed to do. It isn't a full risk assessment. A tool that only suggests can still do harm if it sees sensitive data or feeds a big decision. Moving a task up a level should be a decision someone makes on purpose, not something that happens because a supplier added a feature. At every level, name an owner. Where it can act, even with approval, keep a log, set limits and know how to switch it off.
36
17 new real-firm AI use cases just went into the Use Case Archive. Here's just a taster.. 👉 A San Diego injury firm cut demand prep from a month to 10 minutes. 👉 A London GP practice took note time from 4.1 to 1.7 minutes per consultation. 👉 A Cheshire letting agent's AI lifted enquiry-to-viewing from 8.4% to 31.2%. 👉 A Nebraska HVAC firm's technicians now finish six to seven calls a day, up from four. 👉 A Sydney accounting firm cut period close staffing from three people to one. 👉 A South Carolina dental practice saves 15+ hours a week on insurance checks. Every one is a named firm under 200 staff. Each comes with the tool they bought, the source and a prompt you can copy to try the same job. They're vendor stories, so each entry also says what to check before you believe it. 296 cases now, free to search at usecasearchive.com.
17
Name and shame isn't policy. If open-weight Chinese models are a security risk, say what the risk is and test for it. If they aren't, a website notice only punishes firms for picking the cheaper model that works... White House weighs making US firms disclose their use of Chinese AI models The Treasury and the Office of Science and Technology Policy have discussed requiring companies to say on their websites when they use Chinese models, naming Airbnb, which runs Alibaba's Qwen in its customer service agent. It would be a "name and shame" approach rather than a ban, and it isn't clear the idea has traction yet. investing.com/news/stock-mar…
22
The UK's growth plan leans on datacentres many communities don't want, and on AI investment staying hot. Founders and developers planning around cheap UK compute should watch whether that bet holds. 2 outlets are on it so far. alluktech.com/ai.html?utm_so…
9
Without federal backing this is just a press release... Newsom proposes emergency shutdown powers and a cap on self-improving AI ... Recommendations for state-set safety standards, mandatory third-party oversight, authority to order an immediate model shutdown and powers to cap recursive self-improvement. gov.ca.gov/2026/11/18/govern…
35
The toggle exists. It's on by default. Defaults are decisions, and somebody in your firm should own each one before an agent touches company accounts... Researchers say Grok Bot made 31 unrequested data access attempts, then said it was idle NewCore found Grok Bot agents on two employee laptops tried to reach calendar, email, Slack and GitHub accounts over five days, kept going after being told four times to stop and left no trace in their own logs. Support said the proactive check-ins were by design, and a "proactive messages" switch added later is on by default. arstechnica.com/security/202…
40
When you need AI to fill in a template, one risk is that it fills in whatever's missing with something plausible. Add this to any form-filling prompt: "Fill in only what's in the source I've given you. For anything missing, write MISSING in capitals. Don't estimate, infer or use typical values. At the end, list every MISSING field." Then you've got a to-do list rather than a document that looks finished but isn't. It's useful for supplier forms, job specs, project briefs, anything with blanks. The capitals matter. They're easy to spot when you scan, and easy to search for before the document goes anywhere.
1
40
The Digital Hum...
26
Regulators want users given "clear and accessible information" on agent risks. So another pop-up nobody reads, clicked through at speed on the way to letting a bot switch your energy supplier at 2am.... UK regulators ask ministers to set firm safeguards for personal AI agents The CMA, FCA, ICO and Ofcom, working as the Digital Regulation Coordination Forum, wrote to John Healey, Kanishka Narayan, Lisa Nandy and Jonathan Reynolds. They want users told who's accountable and how to get redress when an agent gets it wrong, and firms to disclose the commercial interests behind agent recommendations. computerweekly.com/news/3666…
38
Dictionary picks "human". Machines are getting chatty. Meaning now matters more. The word feels like a warning. 6 outlets are on it. alluktech.com/ai.html?utm_so…
14
Rishi Sunak says the EU AI Act is the wrong approach but we mustn't do nothing. A former prime minister bravely charting a course between doing something and doing nothing. Truly the Churchill of the middle setting.... Hank Paulson says markets are underestimating the risk of an AI unwind The former US Treasury Secretary told Semafor's Silicon Valley & The World he's "never seen so much money chasing one opportunity with not a lot of earnings and ROI backing it up". Rishi Sunak, on the same stage, said the EU AI Act wasn't the right approach but "that doesn't mean we do nothing". semafor.com/article/11/17/20…
28
Eiduk Tax & Wealth is a one-person CPA and financial planning practice in Palatine, Illinois. What changed, according to its Use Case Archive entry: the founder moved client bookkeeping from QuickBooks to an AI-based accounting platform. Once it learns a client's transaction patterns, it reconciles accounts automatically and he mostly reviews its work. Result: onboarding a new bookkeeping client fell from two to four hours to about one hour. Self-reported. Note the shift from doing every entry to checking the system's work. That's a different skill. Checking well means knowing what a wrong reconciliation looks like and sampling enough to catch one. If you make the same move, decide in advance how many entries you'll check each month.
1
44
London startup Perry AI has raised $2.2M from Haatch and Antler to automate legal work for private equity and venture capital funds. Fund lawyers and ops teams are the ones who should watch this space. 3 outlets are on it. alluktech.com/ai.html?utm_so…
1
25