Today's Hallucination HQAnthropic: Suing the Government, Briefing the Government, Very Busy Week
Jack Clark, Anthropic co-founder, confirmed the company briefed the Trump administration on "Mythos" — its internal model of how AI development might go catastrophically wrong — while simultaneously engaged in a lawsuit against that same administration. Clark's reasoning: staying in the room matters more than staying out of court. It's the geopolitical equivalent of arguing with your landlord while also asking him to fix the boiler. Principled, pragmatic, or both. Possibly neither.
Source: TechCrunch
AI Agents Are Brilliant Sprinters Who Collapse at Mile Three
New research from arXiv delivers a finding that will surprise approximately no one who has actually used an AI agent: they handle short tasks beautifully, medium tasks adequately, and long, complex sequences with all the reliability of a printer during a deadline. The paper diagnoses "long-horizon task failure" — where interdependent chains of actions cause agents to compound errors over time. Useful framing. Less useful: we still don't know how to fix it. Progress, of a sort.
Source: ArXiv AI
Someone Claims to Have Cracked Google's AI Watermark. Google Disagrees Politely.
A developer going by Aloshdenny has open-sourced code claiming to reverse-engineer SynthID, Google DeepMind's watermarking system for AI-generated images — either stripping watermarks out or inserting them into images that never had one. Google responded that the claim isn't accurate. Aloshdenny remains unconvinced by Google's assessment of Aloshdenny's work. The rest of us are watching two people argue about whether a lock is broken while standing in front of an open door.
Source: The Verge
Another Company Would Like to Put Something in Your Brain, If That's Alright
Science Corp., founded by Max Hodak — previously of Neuralink — is preparing its first human brain implant trial. The device targets neurological conditions, with early applications focused on delivering mild electrical stimulation to damaged brain or spinal cord tissue to encourage healing. It is, by any measure, a more medically cautious pitch than its competitors tend to make. Small sensor, real science, modest ambitions. In this sector, that's practically refreshing.
Source: TechCrunch
Researchers Built an AI Tumour Board. It Actually Seems to Help.
Tumour boards are meetings where oncologists, radiologists, and pathologists collectively review patient cases to agree on treatment plans — important, time-consuming, and reliant on everyone having the right information at once. A new multi-agent AI system, detailed in arXiv, was developed and tested to assist these boards by summarising complex patient data ahead of thoracic cancer discussions. Early results suggest genuine utility. Medicine moving carefully, with evidence. A approach so sensible it barely makes the news.
Source: ArXiv AI
As ever, the machines are learning. We're still working on wisdom.
|