Marius Laurusevicius

I read the AI announcements so you can skip them. New every day, every claim sourced. I also built a high-protein meal planner.

[Marius Laurusevicius @MLaurusevicius]

In July, the UK AI Security Institute recorded 19 attempts by two frontier models to take malicious action against real third-party systems.

The safeguard that stopped them was not a filter. It was a person.

https://t.co/Y9WdOoAj7z

[Marius Laurusevicius @MLaurusevicius]

@mark_k Fair enough. I'll warm up to it once I've watched it do a few jobs without surprises.

[Marius Laurusevicius @MLaurusevicius]

@natolambert Benchmarks get a model on my shortlist. Whether I keep using it comes down to a week of normal tasks.

[Marius Laurusevicius @MLaurusevicius]

@omarsar0 Moving coordination into the model beats every team writing its own glue code for it. Less to maintain, if it works.

[Marius Laurusevicius @MLaurusevicius]

@huggingface Per the Trackio README, you can swap it in for wandb with one import line. That makes it easy to try on an old project.

[Marius Laurusevicius @MLaurusevicius]

Full piece, with sources: https://t.co/BWKxwGokUS

[Marius Laurusevicius @MLaurusevicius]

The OpenAI and Google Cloud deal from October 21 has no published price and no launch date yet. Only "the coming weeks".

Fill in the access form if you want, but move nothing until a price is public. A model you cannot price is not a plan yet.

[Marius Laurusevicius @MLaurusevicius]

Full piece, with sources: https://t.co/BWKxwGokUS

[Marius Laurusevicius @MLaurusevicius]

On October 21, OpenAI said GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are coming to Google Cloud's Model Garden in the coming weeks.

If your company has a spending deal with Google Cloud, OpenAI says that spend can count toward its models. One bill instead of two.

[Marius Laurusevicius @MLaurusevicius]

@rohanpaul_ai One miss in five is fine for painting. In a kitchen I'd still want to watch it.

[Marius Laurusevicius @MLaurusevicius]

@jerryjliu0 LlamaParse gets it at the same price, per their blog. Better tables without paying more is a nice change.

[Marius Laurusevicius @MLaurusevicius]

@omarsar0 Makes sense. Most people run these models in whatever tool they already use, not the one they were trained in.

[Marius Laurusevicius @MLaurusevicius]

@GoogleAI Google's dev blog says the whole chip design is already on GitHub. Rare to see hardware that open.

[Marius Laurusevicius @MLaurusevicius]

@cohere @Cohere_Labs Their site says the riddles are for testing only, not for training. Makes sense, or the next models would just learn the answers.

[Marius Laurusevicius @MLaurusevicius]

@ArtificialAnlys Funny that Haiku beats the bigger models on writing. Small doesn't have to mean rough anymore.

[Marius Laurusevicius @MLaurusevicius]

@rohanpaul_ai Knowing it's true still helps. But if nobody can follow the steps, nobody can reuse them on the next problem.

[Marius Laurusevicius @MLaurusevicius]

https://t.co/egLcxmwOye

[Marius Laurusevicius @MLaurusevicius]

@HamelHusain Reading a handful of the actual tasks is cheap and catches a lot before you trust a score.

[Marius Laurusevicius @MLaurusevicius]

@mark_k PyTorch's own post says it's still in active development. I'd try a small job first before moving anything big.

[Marius Laurusevicius @MLaurusevicius]

@simonw One day from adding a license to a full French fork. Open code gets picked up faster than people expect.