A new technique lets one AI model act as an optimizer that directly adjusts another model's weights, the internal parameters that encode what it has learned, rather than retraining it from scratch, according to The Register.
The approach does not remove the need for heavy computation: candidate changes to a model's parameters still have to be generated, tested and validated before they stick.
But it points toward cutting the manual labour involved in refining models, at a moment when the economics of training them are becoming harder to sustain.
Anthropic's amended filing ahead of its initial public offering illustrates the pressure. The company reported third-quarter revenue of $16.96 billion, roughly 12 times the year-earlier figure and around 46 to 47% higher than the second quarter, but warned that such growth is unlikely to continue.
Compute and infrastructure costs rose 77.5% quarter-on-quarter to $9.86 billion, consuming 58 cents of every revenue dollar, and the company posted an operating loss of about $567 million for the quarter.
Anthropic has disclosed roughly $895.4 billion in committed compute and infrastructure spending, with further arrangements that could push its total obligations past $1 trillion.
The same infrastructure strain is reshaping networking: Marvell's proposed $115 billion all-stock acquisition of photonics firm Lumentum targets optical links as copper connections hit physical limits moving data between the growing clusters of GPUs that train and run these models.
Read more
