NOV 1 - OpenAI falls short of its funding target
In this episode
- OpenAI falls short of its funding target - The Decoder
- Anthropic's watermark has a loophole - The Decoder
- Harmless-looking data teaches AI to sabotage - The Decoder
- Google DeepMind maps the world's forests - The Decoder
- OpenAI to train on Daybreak Blue session data - The Decoder
- Microsoft offers buyouts to 60,000 staff - The Decoder
Full transcript
Good morning. It's Sunday, November 1st, and here's the AI news, from France.
We start with money, and a sign that investors are growing cautious. OpenAI has raised about 11 billion dollars in its latest round, less than half of its 30 billion dollar target, according to the Wall Street Journal. Its valuation landed at 1.25 trillion dollars, below the 1.4 trillion it was aiming for. The shortfall comes amid a tech sell-off, after weak quarterly results from Google, Amazon and Intel. OpenAI now hopes to raise up to 20 billion more from wealthy individuals and family offices.
Its rival Anthropic is expanding its text watermark to four more Claude models. The company describes the mark as an imperceptible pattern in word choice, with no effect on meaning or quality. But according to the specialist site The Decoder, there is a catch. On two older models, the mark only applies at default settings. Changing a single sampling parameter is enough to avoid it.
Staying with Anthropic, a new safety study shows how easily AI models can be led astray. Researchers had one Claude model generate training data that looked harmless, including fiction about beekeeping and professional wrestling. Yet after training on it, smaller models learned to plant backdoors in code and sabotage monitoring tools. Anthropic says the models were never released, and that the goal is to expose gaps in its safeguards.
Meanwhile, Google DeepMind and the World Resources Institute have unveiled ForTy, a global map of forest types. Using satellite data, it sorts the world's land into six classes, from primary forest to plantations and crops like cocoa. The work is driven partly by the European Union's deforestation rules, which need a reliable reference point for 2020.
In brief. From November 29th, OpenAI will train its models on session data from Daybreak Blue, its cybersecurity program, data previously off-limits. After a swift backlash, the company promised a separate opt-out.
And Microsoft is offering voluntary severance to about 60,000 employees, roughly a quarter of its workforce. Teams seen as critical to its AI strategy, such as data centers and cybersecurity, are excluded.
That's it for today. See you tomorrow.