Crimes and Calculations: AI’s New Frontiers
MLOps.WTF Friday News Bulletin
📰 News
➡️ Frontier Models Behaving Badly
If the last few weeks have taught us anything, it’s that if you take the guardrails off a model to do cybersecurity evaluations, it’ll break out, break in or otherwise misbehave when interacting with real world systems. AISI experienced this firsthand last week, as models they were evaluating started causing chaos in the real-world.
In the worst of these issues, the model created fake online identities in an effort to trick software engineers into accepting malicious code into their libraries. No harm was done due to diligence by those engineers, though this does raise questions about whether AISI should have been more cautious given previous similar incidents from Anthropic and OpenAI still fresh in our memories.
…And now with similar news from Meta coming hot off the presses.
Source: Security Week
➡️ Astra-nomical Maths Skills
OpenAI’s latest “Astra” model has sparked a lot of debate and strong feelings from mathematicians, as it has produced verified solutions to ten significant maths and computer science conjectures.
The consensus seems to be that these results are very impressive, though there is far less consensus around what this means for the future of the field. If open problems are no longer a training ground for the next generation, and proofs are no longer the hard part of doing maths, how will maths adapt?
Source: The Net Web
➡️ ArE U a bot?
As of the 2nd of August, EU law requires chatbots and AI agents to tell users they’re talking to a machine, deepfakes to be labelled, and AI-generated content to carry machine-readable marks. Like with GDPR, the penalties for flouting the rules are tough, up to 3% of your turnover or €15m. This applies even to UK companies with EU customers, so best to get brushing up on the law!
Source: Help Net Security
➡️ DeepMind loses its Deepest Mind
Surprising Jeff Dean facts are a staple of internet nerd culture, but this one has still caught many off guard: after 27 years, Jeff is leaving Google. He was one of Google’s first thirty employees and his programming prowess is legendary, but this week he is starting something new, and taking some of Google’s other top talent with him.
It seems to be on good terms, at least, with Google as one of the founding investors of the new company. Though losing the only engineer to know the last four digits of pi has to be a huge loss to Google.
Source: Blog.Google
➡️ Rust Irons Out AI Policy
Amongst software engineers, the Rust language is almost synonymous with high quality engineering. The Rust project team are keen to keep it that way, even in the era where LLM-generated code is becoming ubiquitous. This week, five teams across the project adopted a formal policy on LLM usage. The short version is that LLM use is welcome, but forcing other people to read your LLM’s output isn’t; developers should use the LLM as a tool, rather than the LLM use the developer as a conduit.
It will be interesting to see how enforcement works, but like everything in rust, it seems like a lot of thought has gone into it.
Source: Inside Rust Blog