A new test says AI is getting much better at running experiments on its own, and an Anthropic co-founder now sees self-improving AI arriving by 2027.
Jack Clark, a co-founder of Anthropic, now thinks AI that helps build its own successors could arrive as early as 2027. He had expected 2028. What changed his mind is a new test called TasteVal, which suggests AI is getting good at designing and judging experiments. The test has limits, and its own authors say it may make progress look faster than it is.
TasteVal comes from P-Zero Research. The lab uses the term "research taste" for the skill of choosing the right experiments and drawing the right conclusions from them. TasteVal checks only the last two of three parts. It does not test whether an AI can pick which research question is worth pursuing. That choice is already made in each of its eight tasks.
The tasks come from AI model development, such as selecting training data and defending against adversarial prompts. The model decides which experiment to run next and uses the results to guide its next step.
P-Zero Research turned research taste into a simple ratio: how much computing power a model needs to match human experts. The human group was 24 experts who had recently worked at places like OpenAI, Google DeepMind, Nvidia and Microsoft Research. The best human attempt on each task set the bar.
One benchmark cannot show a lasting trend. The authors admit their results may overstate how quickly research taste is improving. The tasks can be checked quickly and cheaply, which is the kind of work AI labs find easiest to optimize for. The human group also had no top-tier researchers.
Clark's forecast carries its own bias, since his company makes the top-scoring model. In a May essay he put the odds of an AI that trains a stronger successor on its own at about 30 percent by 2027 and 60 percent by the end of 2028.
Critics have also pushed back. Researcher Herbie Bradley argued that skills like creativity and long-term research planning were nowhere in sight. A Fudan University study found AI agents contributed up to 55 percent of methodology proposals, but humans made over 85 percent of final decisions.
Nothing here says AI is already improving itself. But if AI starts doing real research work, products could change faster. For anyone who relies on AI tools or vendors, it is worth watching how quickly those tools update and how their makers describe the changes.
A new test shows AI getting fast at running experiments, but it covers only part of research, so treat the 2027 forecast as a signal to watch, not a fact.
Lawmakers in both parties want Nvidia's CEO to answer questions about a reported tie to an alleged Chinese intelligence asset, and chip rules could shift as a result.
Read more: Lawmakers from both parties want Nvidia's CEO to answer questionsDeepSeek says 1.54 million people use its AI agent platform each day. If you run a business, cheaper AI helpers may soon be offered to you.
Read more: DeepSeek says its AI helper platform now has 1.54 million daily usersOpenAI has about $11 billion in hand at a $1.25 trillion value, below its first ask. Microsoft and NVIDIA shareholders have reason to watch what comes next.
Read more: OpenAI raises $11 billion, but at a lower price than it wanted
