In an X thread, Google DeepMind researcher Andreas Kirsch claims open-source models now unlock applications impossible with closed systems two years ago and projects they will soon reach the level of Anthropic's Mythos line; research engineer Florian Brand replies that such forecasts need concrete evidence rather than optimistic framing.
Not sure, but open-source is close behind and soon enough it will be as good as Mythos is now
The better question to predict the future is: are there examples that open source AI enables right now that weren't possible with closed models two years ago?
And the answer is of course yes ☺️
@BlackHC Are there any examples where open source AI enabled something that wasn’t possible with the closed models?
No benchmarks, datasets, or named use-case examples appear in the exchange to support the new-capabilities claim or the near-term parity projection, so the discussion rests on assertions whose timing remains unspecified.
The back-and-forth highlights the recurring tension between forward-looking statements on open-source progress and demands for verifiable deltas, leaving participants to weigh the same projections without fresh technical anchors.
Users reacted to analyses claiming closed AI models enable hacking and scams, with some praising the coverage while others called it insufficiently harsh on Anthropic and warned that concentrated funding is harmful.
New blog.
I looked into the actual evidence and what models where used by bad actors to see whether closed models are safer.
Turns out: Nope, they are used to hack, misinform and scam. There is one exception, though.
Link in replies.
https://florianbrand.com/posts/open-model-safety
New blog.
I looked into the actual evidence and what models where used by bad actors to see whether closed models are safer.
Turns out: Nope, they are used to hack, misinform and scam. There is one exception, though.
Link in replies.
Any reasonable analysis of the past five years makes it abundantly clear that major AI companies are making the world a worse place and open source AI isn’t. So they distract you from the actual harm they’re doing in the world with hypothetical fearmongering about the future.
New blog.
I looked into the actual evidence and what models where used by bad actors to see whether closed models are safer.
Turns out: Nope, they are used to hack, misinform and scam. There is one exception, though.
Link in replies.
@BlackHC I prefer evidence over vibes
Not sure, but open-source is close behind and soon enough it will be as good as Mythos is now
The better question to predict the future is: are there examples that open source AI enables right now that weren't possible with closed models two years ago?
And the answer is of course yes ☺️
The confounder as you point out is that closed models are more capable and easier to use hence there is a little incentive to use open models or train/tune your own
I think your point that open models are constantly improving and surpassing what closed models used to be able to do is more important
In the future, if it gets too difficult to use closed models, open models will then have the capabilities that were required for these misuses (but there will be less monitoring etc bc they will run locally)
New blog.
I looked into the actual evidence and what models where used by bad actors to see whether closed models are safer.
Turns out: Nope, they are used to hack, misinform and scam. There is one exception, though.
Link in replies.
@BlackHC For this to be wrong, it requires coordination by Anthropic, OpenAI, GDM and xAI to close down the same capabilities and make them all similarly hard.
I think this is unlikely given the current landscape, the OpenAI reactions to the nerfed capabilities were telling
The confounder as you point out is that closed models are more capable and easier to use hence there is a little incentive to use open models or train/tune your own
I think your point that open models are constantly improving and surpassing what closed models used to be able to do is more important
In the future, if it gets too difficult to use closed models, open models will then have the capabilities that were required for these misuses (but there will be less monitoring etc bc they will run locally)
@BlackHC Looking at the actual evidence at hand and it’s looking awful for closed models in terms of safety. Closed guardrails don’t seem to work at all. Happy to discuss this later once i presented the facts.
yeah, looking further into this and it seems like open models are safer. blog about this later
@xeophon Nice read! 👏
@xeophon @diari_cc So when closed models get locked down, what will become the easiest route then?
Isn't this a confounder for a historical analysis?
Maybe there is no need for using open-source models *just yet*?
@BlancheMinerva
New blog.
I looked into the actual evidence and what models where used by bad actors to see whether closed models are safer.
Turns out: Nope, they are used to hack, misinform and scam. There is one exception, though.
Link in replies.
@xeophon Was this post before or after Mythos cause you missed a great one. "The Mythos of unsafe..."
@xeophon Also given past events, I could imagine a reporting requirement in the future (similar to someone telling their doctor or therapist when they seem to be seriously considering harming themselves or others). This would also shift use towards local models
@xeophon @BlackHC Maybe I am wrong but surely with open source models you have a better chance to remove safety guardrails through fine tuning.. as per the papers below
https://www.usenix.org/conference/usenixsecurity25/presentation/song-minkyoo
https://openreview.net/forum?id=hTEGyKf0dZ&utm_source=chatgpt.com
@BlackHC @diari_cc Well my argument is wrong when the closed frontier coordinates a collective pause and halts AI development, yes. Doesn’t seem to be the case, though.
@xeophon @diari_cc So when closed models get locked down, what will become the easiest route then?
Isn't this a confounder for a historical analysis?
Maybe there is no need for using open-source models *just yet*?
@maximelabonne Thank you!
@xeophon @N8Programs
The confounder as you point out is that closed models are more capable and easier to use hence there is a little incentive to use open models or train/tune your own
I think your point that open models are constantly improving and surpassing what closed models used to be able to do is more important
In the future, if it gets too difficult to use closed models, open models will then have the capabilities that were required for these misuses (but there will be less monitoring etc bc they will run locally)
@xeophon Not harsh enough
I should write one that explains why ant is evil
@xeophon That's not vibes. That was a pretty good argument (better than yours imo)
@xeophon I'd be willing to bet that governments will step in eventually if capabilities keep increasing in this trajectory
The USG already requested a delay of the release of Mythos iirc
@BlancheMinerva I think that's a pretty fair criticism. The problem is when two startups get the majority of all funding it is harmful in and of itself. Of course if they behave like bad actors it's certainly a dangerous use of capital.
In an X thread, Google DeepMind researcher Andreas Kirsch claims open-source models now unlock applications impossible with closed systems two years ago and projects they will soon reach the level of Anthropic's Mythos line; research engineer Florian Brand replies that such forecasts need concrete evidence rather than optimistic framing.
Not sure, but open-source is close behind and soon enough it will be as good as Mythos is now
The better question to predict the future is: are there examples that open source AI enables right now that weren't possible with closed models two years ago?
And the answer is of course yes ☺️
@BlackHC Are there any examples where open source AI enabled something that wasn’t possible with the closed models?