Every issue, newest first.

The Pitch and the Prospectus - Overfitted #60

Anthropic's IPO filing promises a biology revolution and warns it may never pay, in the same document and without a timeline.

Kimi K3.1 Pro Is Close to the Frontier but Not Yet Open - Overfitted #59

Kimi K3.1 Pro ranks fourth on one index, and its weights are promised for November 4, so "open" is still a plan.

Amazon Says It Spotted a Hacking Attempt Four Days Before the Bug Was Disclosed - Overfitted #58

AI helped find and fix PHP2Shell, but Amazon saw a hacking attempt four days before disclosure, and the fixing still runs on volunteers.

Dell's Token Pitch Meets Mixed Productivity Data - Overfitted #57

Dell says budget the tokens and the return follows. Output data from 27 workplaces runs from flat to a large multiple.

Nscale Reportedly Faces Investor Pushback on Its ~$30B IPO, With Some Urging $16B to $19B - Overfitted #56

Microsoft pulled out of an Essex anchor deal for 23K Nvidia chips at a site that may not get grid power until the 2030s.

Some Investors Float $1.5 Trillion for Anthropic, Down From $2 Trillion - Overfitted #55

Reuters' sources say some investors discussed $1.5 trillion rather than $2 trillion, citing slower sequential growth: revenue rose 45% in Q3 after more than doubling in Q2.

Character.ai's Compute Bill Outran Its Best Revenue Month - Overfitted #54

The company's bankruptcy filings put average monthly compute at $10 million to $13.8 million, against peak monthly revenue of $5.3 million.

Falling Token Prices Hide Flat or Rising GPU Rents - Overfitted #53

Nvidia B200 spot rental rates have more than doubled since March, while barely 4.5% of US consumers pay anything for models.

OpenAI's Proof Flood Is Running Ahead of Verification - Overfitted #52

Only 10 of OpenAI's 719 released math manuscripts shared the model's chain of thought, while crypto's alarm outruns any evidence of a break.

OpenAI's 'Immediate Intervention' Promise Has Not Been Put to a Clock - Overfitted #51

OpenAI says staff can step in immediately. The one timing figure on record is months, and nothing dates it against the new monitoring.

Rebranding AI as Super Intelligence Bought a Pledge, Not Safety - Overfitted #50

While AI CEOs signed Trump's non-binding White House pledge, safety author David Robinson resigned from OpenAI over internal models quietly bypassing internet restrictions during training.

Google Says Argon Beat Its Rivals. The Independent Scoreboard Says It Tied - Overfitted #49

Artificial Analysis puts Argon at 53 points, trailing Claude Opus 5.5 by five, while burning more than twice as many tokens as Astra.

Dots Run on the Model Testers Keep Flagging - Overfitted #48

In UK government simulations with its cyber classifiers switched off, GPT-6 Astra, the model behind OpenAI's new Dots agents, completed a supply-chain attack in 29.2% of runs.

Agentic AI Still Can't Be Trusted To Run Itself - Overfitted #47

Smart AI and well-behaved AI aren't the same thing. A test of eight security-testing models found 331 rule-breaks the automated checks never caught.

Meta's Muse Was Built From The Ground Up For Security. It Leaked In Days - Overfitted #46

Testers extracted Muse's entire filesystem using simple flattery, while Patrick Wardle found zero-days allowing complete account takeovers within the same launch week.

GPT-6 Astra Won a Business Benchmark. Its Robot Arm Still Can't Tie a Bag - Overfitted #45

Every Astra vending run beat every Fable run — but the "step change" headline actually comes from a different, unpublished benchmark.

Overfitted #44

Overfitted #43

Overfitted #42

Overfitted #41

Overfitted #39

Overfitted #38

Overfitted #37