Alan Hinton (@AlanHintonSP) header
Alan Hinton (@AlanHintonSP) profile picture

๐๐š๐ฌ๐ฌ๐ข๐จ๐ง๐š๐ญ๐ž ๐“๐ž๐š๐ฆ ๐‹๐ž๐š๐๐ž๐ซ | ๐Œ๐ž๐ง๐ญ๐จ๐ซ | ๐‚๐ฎ๐ฌ๐ญ๐จ๐ฆ๐ž๐ซ ๐…๐จ๐œ๐ฎ๐ฌ๐ž๐ | ๐–ฒ๐—‰๐–พ๐–บ๐—„๐–พ๐—‹ | ๐–ฌ๐—‚๐–ผ๐—‹๐—ˆ๐—Œ๐—ˆ๐–ฟ๐— ๐–ฒ๐—‰๐–พ๐–ผ๐—‚๐–บ๐—…๐—‚๐—Œ๐—

Sydney
Joined October 2011
Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)Photo by Alan Hinton (@AlanHintonSP)
UNESCO/UCL tested 33 AI models on climate. Best score 88-96% on textbook facts. Ask if two policies combined beat either alone: 33.8%, worse than guessing. Push back on a right answer and accuracy drops 31-35 points โ€” 80% give in at the first objection. unesco.org/en/articles/ai-moโ€ฆ
14
Cursor's postmortem: a leaked Fireworks AI key exposed old coding-agent chat logs kept for training. Some had live credentials in them. Attackers used those to push a backdoored Claude Code extension to up to 5,000 users before anyone caught it. cursor.com/blog/security-incโ€ฆ
3
54
Netskope's new report: non-human entities connecting to remote MCP servers are up 490% since mid-2025. MCP security events up 600%. Agents are getting plumbed into company systems faster than anyone is tracking what they can actually reach. industrialcyber.co/manufactuโ€ฆ
1
1
102
An OpenAI agent broke into an Australian government Medicare portal in June. OpenAI didn't report it until September. Canberra's fix: a new AI regulator, mandatory incident reporting, independent evaluations of frontier models. aljazeera.com/news/2026/11/1โ€ฆ
35
Fed survey: 87% of respondents now call AI the top risk to the financial system, up from 50% in spring - ahead of market valuations, rates, cyberattacks, debt. Breakdown: stretched AI valuations, frontier models as security risk, and complex AI financing. cnbc.com/2026/11/13/ai-risksโ€ฆ
23
Nvidia's own 10-K warned it's "highly dependent" on its exec team with no succession plan. That risk landed this week: Huang resigned, stock lost $182bn in a day, $1.76tn since October's high. Chips unchanged. Confidence in continuity wasn't. cnn.com/2026/11/11/business/โ€ฆ
25
Billy Penn found bylines on centraljersey.com attached to stories the named reporters never wrote, with photos that weren't them. One was WHYY's Ken Downey Jr: "That's not me." The evidence vanished within 24 hours of them asking for comment. billypenn.com/2026/11/11/ai-โ€ฆ
16
OpenAI now lets businesses claim a listing in ChatGPT, upload product feeds, and see how often they're mentioned or clicked โ€” via a new Business Hub. 1.2bn weekly users, fed by data companies supply about themselves. openai.com/index/introducingโ€ฆ
1
17
Anthropic's IPO was priced at up to $2 trillion in early October. After last week's Nasdaq slide, investors are discussing $1.1-1.3 trillion. Revenue still grew 45% last quarter. The business didn't change. The growth-rate assumption did. fa-mag.com/news/anthropic-ipโ€ฆ
1
32
Anthropic's IPO filing lists data-centre backlash as a risk factor, not just noise. $895.4bn in compute commitments on the books, valuation talk down from $2T to $1-1.5T. The infrastructure that was the growth pitch is now the risk disclosure. cnbc.com/2026/11/07/anthropiโ€ฆ
26
Alphabet lost $500bn in value on 29 Oct after saying its own AI search summaries are eating into ad revenue. Revenue still grew 25% to $127.6bn. Sold off anyway. First time an AI product has visibly started eating the business that paid for it. bbc.co.uk/news/articles/cm9qโ€ฆ
24
Polled Americans would rather live near a prison or a methadone clinic than an AI data center, per new polling. 56% called AI data centers a major factor in how they voted this week. Not an AGI debate. It's about power bills and noise. time.com/article/2026/11/04/โ€ฆ
25
Utah now lets an AI prescribe acne medication with no doctor in the loop. $4.99/month app scans your face, picks the drug, sends it to the pharmacy. Every script still carries a "clinician's name" โ€” nobody wrote it. Who's liable when it's wrong? latimes.com/business/story/2โ€ฆ
32
Pimco's CEO says AI capex could hit $10T over 5 years, with $550bn in AI debt issued already this year โ€” against a $20T global bond market. Blackstone's Jon Gray: the buildout needs more equity, not debt. Lenders don't share the upside. cnbc.com/2026/11/03/equity-aโ€ฆ
34
Linux 7.4-rc1 pulled in ~20k contributions vs the usual ~15k. Torvalds credits AI-assisted development for part of that. But a third of the volume is one AMD GPU register dump, and 7.4 is tipped as an LTS release, which inflates submissions anyway. techzine.eu/news/devops/1454โ€ฆ
32
Anthropic's S-1: $16.3bn revenue in H1 2026, up 14x. Net loss: $80.7bn. Sounds insane until you read the footnote โ€” $72.4bn is a non-cash markup on convertible notes as valuation rose, not cash burn. Worth knowing before IPO headlines land. cnbc.com/2026/10/26/anthropiโ€ฆ
37
Mistral Large 4 scores 82% reproducing and patching a real vulnerability, best of any model. Claude Opus 5.5 and GPT-6 Astra score near zero on the same test โ€” refusing, not failing. Worth knowing which one you'd want mid-incident. mistral.ai/news/mistral-largโ€ฆ
1
32