Skip to content
DevMeme
6471 of 7590
AI ML Post #7095 · source on Telegram

EU AI Act Enforcement Depicted as Police Raid for Fine-Tuning Llama 2

Description

A screenshot from X.com showing two posts from xlr8harder (@xlr8harder). The first post (10h ago) reads: 'when you forget to notify brussels you are finetuning llama 2' with a dramatic fish-eye lens photo showing what appears to be a police raid/SWAT team entry into a building at night. The second post (13h ago) references: 'The first deadline for EU AI act reporting passed in August, and all models over 10^23 flops must now be formally reported on to a regulatory a...' alongside a pie chart showing 'Continents that have an AI Regulation' with 'The EU' (large blue circle) vs 'Others' (tiny dot). The joke satirizes EU over-regulation of AI, suggesting that even fine-tuning an open-source model like Llama 2 could trigger a regulatory enforcement action

Comments

31
Anonymous ★ Top Pick EU AI Act compliance: where fine-tuning Llama 2 on your gaming PC requires more paperwork than actually building a nuclear reactor -- at least the reactor has a clear regulatory pathway
  1. Anonymous ★ Top Pick

    EU AI Act compliance: where fine-tuning Llama 2 on your gaming PC requires more paperwork than actually building a nuclear reactor -- at least the reactor has a clear regulatory pathway

  2. Anonymous

    Nothing says "model governance" like discovering your change-management approval queue now includes battering rams and a FLOP counter

  3. Anonymous

    The real horror isn't the 10^23 FLOPS threshold - it's explaining to the compliance team why your 'small experimental fine-tune' accidentally spawned 47 checkpoint files across three regions, each technically requiring separate notification because someone forgot to set FSDP_STATE_DICT_TYPE=FULL_STATE_DICT

  4. Anonymous

    The EU AI Act's 10^23 FLOPS threshold is basically the regulatory equivalent of 'if your model can pass the Turing test, it better pass our compliance test first.' Turns out the real alignment problem isn't getting AI to follow human values - it's getting ML engineers to follow Brussels' paperwork deadlines. At least when your model hallucinates, you can blame the training data; when you hallucinate that compliance is optional, you get a fisheye view of regulatory enforcement at your door

  5. Anonymous

    Finetuning Llama 2 to self-report under EU AI Act: where compliance FLOPs eclipse inference budgets, proving regulators scale worse than any model

  6. Anonymous

    Nothing says Responsible AI like measuring risk in training FLOPs - my LoRA tune on a single 3090 triggered more paperwork than gradients

  7. Anonymous

    Apparently kubectl apply -f notice.yaml is the new hyperparameter - skip it and your training run’s observability downgrades to door-peephole monitoring

  8. アレックス 11mo

    Doesn’t disk encryption and/or paying Jensen for cloud time render this completely unenforceable?

    1. @niko2048 11mo

      a lot of stuff in the EU does, it's more about formality I guess

  9. Max 11mo

    These are details politicians don't care about. If all rules were enforced we'd live a miserable life :D Also, I think flops are the wrong measure to measure model complexity; number of weights would be better. FLOPS isn't even constant, as it scales with the input size o.O

    1. @niko2048 11mo

      yeah my thoughts too

    2. @SamsonovAnton 11mo

      My guess is that FLOPS is mistaken for FLOPs often, i. e. they meant just a niumber of FP operations in total — not per second; more precisely, they actually meant just the amount of FP data rather than the number of operations on it.

      1. @unikol 11mo

        Probably the funniest part of this law is that if we find a reliable way to train the models in integer parameter space the arbitrary FLOPs limit becomes completely useless.

        1. _ 11mo

          Aren't some local / tiny models already using fixed point ? Then they're already using interger space

          1. @unikol 11mo

            The limit is set for compute spent on training a model, not inference

  10. @niko2048 11mo

    Still bummed about Chat control, hopefully that fails miserably

    1. @Johnny_bit 11mo

      I've contacted ALL of my country reps to vote against chat controll shite. Hopefuly more people across whole eu will do the same.

      1. dev_meme 11mo

        Did I actually make a post with link how to contact MEPs? I don’t remember somehow I might have missed and not posted it yet

  11. @Johnny_bit 11mo

    also - eu is overproducing regulations and bureaucracy.

    1. @itimonin 11mo

      commonly known as crazy printer(at least in our place)

  12. @callofvoid0 11mo

    what is flops?

    1. _ 11mo

      Floating point operations per second https://en.wikipedia.org/wiki/Floating_point_operations_per_second

      1. @RiedleroD 11mo

        flops as a way to measure how good a GPU is is like measuring a car based on horsepower alone. the simple fact that reality is a lot more complicated than that flies over the head of consumers

        1. @RiedleroD 11mo

          150TFlops! ok but if it only has 2GB of VRAM it's trash

          1. @RiedleroD 11mo

            same way a car is trash if it has 600 horsepower but weighs 5 tons

        2. dev_meme 11mo

          Consumers? You meant bureaucrats, right?

          1. @RiedleroD 11mo

            anyone who isn't an expert in the field

  13. @SamsonovAnton 11mo

    When AI will finaly rise to power,, every human with IQ > 23 should be reported.

  14. dev_meme 11mo

    No, but you can contact office of MEP

  15. dev_meme 11mo

    The same way you can contact office of senator from your state in US

    1. dev_meme 11mo

      And some of them may reply even personally (especially when it’s aligned with their agenda and will look good in press)

Use J and K for navigation