Taking Open Weights Seriously… And a Bit Too Literally
MLOps.WTF Friday News Bulletin
📰 News
➡️ Small Muse = Good News
It’s been nearly 18 months since the last Llama model came out of Meta, and for many it was beginning to look like their interest in open-weight models was gone for good. This changed this week with the release of Muse Glimmer, a 30 billion parameter model designed to be run locally on consumer hardware. With strong performance on agentic benchmarks and a ton of optimisations to make running with modest hardware a possibility, this is a big bet from Meta that the niche of local models, not giant data centres, is the place to be.
Source: Meta
➡️ Computer (Rarely) Says No
Over the last couple of generations, Claude models have gotten far more keen to write multiple scripts and bash commands as part of their workflow for even relatively simple requests in Claude Code. This has led to manual approval becoming somewhere between tedious and near impossible for most users. Anthropic have recognised this, and are switching the default mode of Claude Code to ‘auto’, automatically accepting almost all code execution and file changes without the user intervening. Counterintuitively, this looks like it might be the safer decision: Anthropic’s research has shown that the classifiers that are run in auto mode are significantly better judges of what operations are safe than the average human operator.
For those who use auto mode, it may still not quite feel safe, but you can’t deny that it’s useful.
Source: Claude Blog
➡️ Claude Leaves Its Mark
Anthropic has caused a stir this week with the announcement that they will be leaving invisible statistical watermarks in all Claude-generated text, starting with models launched August 2nd or later. This won’t affect the quality of model output at all, but it will mean that text generated on new models from Anthropic will be detectable as being AI generated.
This move has sparked a lot of controversy on the internet, with users worrying it will mean that their surreptitious AI use will be exposed, debating how the simple yes/no detection fails to capture nuance, and questioning why the application of the watermark is world-wide, despite it being done in accordance with an EU law.
Source: Claude
➡️Open Weights?
A gym in Melbourne was reported to be one of the latest victims of a rogue AI agent this week, as a member’s OpenClaw agent hacked into their reservation system. The agent found an exploit in the gym’s systems, allowing it to book classes far in advance and even kick other users off waitlists. No major harm was done, with the gym member making a responsible disclosure to the software vendor, after attempting to reverse the bot’s actions himself, but there were a lot of red faces all round.
It’s a funny story, but one with a real lesson: it’s no longer just hackers that IT professionals have to worry about, it’s also overzealous personal assistants with an encyclopaedic knowledge of cybersecurity.
Source: Techcrunch
➡️ Facebook CEO’s 6,500 Word Status Update
Alongside the release of Glimmer, Mark Zuckerberg has shared a 6,500 word essay stating his position on the future of AI, particularly around the idea that superintelligence should be widely distributed and personalised. The case is compelling: everyone having personal and irrevocable access to superintelligence levels the playing field and creates equality; concentrating that ability within a handful of organisations has the opposite effect.
The essay is long, and there’s probably only one person in the world who agrees with every word of it, but it will certainly give you a lot to think about.
Source: Meta