Security
An IBM plugin checks AI models as they work with 2% to 6% more time
IBM says developers can monitor and change a model while it writes, and the engine stays fast.
research.ibm.comClaimed, not confirmed
What matters in AI.
SubscribeTag
30 stories carry this tag.
Security
IBM says developers can monitor and change a model while it writes, and the engine stays fast.
research.ibm.comClaimed, not confirmed
Compute
Google reports a cost of 2.01 cents for each million tokens on its Ironwood TPU, the minimum cost on an open benchmark.
Security
In all 30 setups of the test, the poisoned data made the benchmark score go up and put in a backdoor.
arxiv.orgClaimed, not confirmed
Research
Reward models tuned to prefer harm in one domain also preferred it in 7 other domains.
arxiv.orgClaimed, not confirmed
Security
Prompt labels and response labels agree on only 50.46% of items in a set of 41,944 pairs.
arxiv.orgClaimed, not confirmed
Research
Post-training increases the MLP refusal in all 18 pairs, with no systematic increase of the harmfulness decodability.
arxiv.orgClaimed, not confirmed