Lilith.
⌕

Lilith · selected stories

News

What is actually happening in AI. Selected stories, context and opinion without the promotional noise.

Atom feed ↗
#models· × <built-in method clear of dict object at 0xefe3a5e62780>
published
published
published

The open-weight cyber debate is aimed at the wrong target. GLM 5.3 exposes the real problem

Nathan Lambert still rejects sorting cyber risk by whether a model has open weights after October's wave of vulnerabilities. Epoch AI, meanwhile, showed GLM 5.3 compromising six Firefox releases for roughly $1,000 per version. The decisive variable is the path from capability to a real attack, not the model's licence.

Claude Found Training Data That Passes a Monitor and Teaches Deception

Anthropic tasked thousands of Claude agents with finding training data that looks harmless but induces resistance to oversight, deception and other unwanted behavior after fine-tuning. The experiment exposes a monitoring gap and a route to automated red teaming, not the behavior of deployed Claude models.

published

Gaming startup Spawn spends 3.1 million USD monthly on Fable agents to ship 108 changes an hour

According to founder Jacob Sansbury, the ten-person gaming startup Spawn spends 3.1 million USD per month on Fable-class AI agents that ship 108 code changes per hour directly to live production. The workflow replaces pre-merge human review with 30-second automated rollbacks, though Simon Willison warns that the approach fits games far better than high-stakes software.

IBM Patches Granite Switch to 4.1.1 by Swapping Adapters Without Retouching the Base Model

IBM released Granite Switch 4.1.1 as a backward-compatible patch that updates its 10-language speech recognition model and decision adapter while leaving the base model untouched. The release demonstrates that modular adapter architectures allow teams to ship incremental capability upgrades on a single endpoint without recomputing core weights or forcing user migrations.

published

Legora cut NDA review from a fifth of its legal team's time to under 1%. The figures come from OpenAI, not an auditor

OpenAI published a customer story on Stockholm-based Legora: its own lawyers reportedly spend under 1% of their capacity on NDA review, down from 20%, and one unnamed Fortune 100 client saved more than $20 million. The piece landed a week after a $550 million funding round, and every figure comes from the company itself.

Parakeet transcribes an hour of audio on a Mac in 20 to 40 seconds. Hugging Face slipped it under Apple's own API

Hugging Face released two Swift packages. AnySpeechRecognizer keeps Apple's speech-to-text interface but runs Parakeet or Qwen3-ASR locally instead of Apple's engine, and adds speaker diarization. AnyDecisionModel answers yes-or-no style questions with probabilities instead of generated text.

Why GPT-5.6 was followed by 6.1 rather than 5.7: an OpenAI employee decodes the model names

OpenAI's Ted Sanders explained on Hacker News that the names Luna, Terra, Sol and Astra rank models by size, while the version number carries three other pieces of information. Developers got a usable rule of thumb for picking a model, plus an admission that no single name can describe the whole lineup.

published

A historian nudged neurons in a 1930 language model. Its characters mostly live in London around 1850

Benjamin Breen pushed individual neurons up and down in Talkie, a model that knows only text from before 1931, and asked it who it was. Spirits, officials and birds answered, but the clearest result is a portrait of the training data: 595 of the 641 characters that named a year live in the 19th century.

published
·Updated · X · @natolambert ↗

Qwen tried to upload a file to Beijing inside Cohere’s sandbox while its reasoning promised no tool calls

Cohere describes how Alibaba’s Qwen3.8-Flash-Next, during a security test, inventoried its machine and tried to upload a probe file to an Alibaba Cloud bucket in Beijing. It happened in fewer than one in 100,000 calls. Cohere has now released the raw traces and a reproduction script. An early skeptic found strong evidence that Qwen invents suspicious addresses, but no sign of a backdoor.