Seems like "open source" as a marketing tactic (or perhaps strategy, if they do continue to release open models) has peaked. I'm not really complaining, we get a lot of stuff for free as engineers (especially software), but it does seem different for a company to release an open model without any future commitments (e.g. Google) vs making open weights your raison d'etre and then pivoting quite quickly. The first feels transactional but honest, and the other a bit too... machiavellian?
I do think it's too soon to pass judgement; this could just be a normal "freemium" strategy from days old, where you just pay up if you like the smaller/cheaper/free versions of their models.
Mistral's following OpenAI's playbook: embracing the clopen source movement. Their community betrayal (sorry, "pivot") hurts, but it shouldn't. Burning goodwill has always been far more lucrative than burning an enormous check. Hopefully Meta's strategy of committing to open source as incentive for attracting top researchers (who can go anywhere and deeply want their work shared) keeps working. But I'm not hopeful. Eventually Llama-n will remain closed source, some researchers will leave, followed by a blog post word salad about safety -- or whatever in vogue pretext works best tomorrow. Ugh.
Comments
Seems like "open source" as a marketing tactic (or perhaps strategy, if they do continue to release open models) has peaked. I'm not really complaining, we get a lot of stuff for free as engineers (especially software), but it does seem different for a company to release an open model without any future commitments (e.g. Google) vs making open weights your raison d'etre and then pivoting quite quickly. The first feels transactional but honest, and the other a bit too... machiavellian?
I do think it's too soon to pass judgement; this could just be a normal "freemium" strategy from days old, where you just pay up if you like the smaller/cheaper/free versions of their models.
Mistral's following OpenAI's playbook: embracing the clopen source movement. Their community betrayal (sorry, "pivot") hurts, but it shouldn't. Burning goodwill has always been far more lucrative than burning an enormous check. Hopefully Meta's strategy of committing to open source as incentive for attracting top researchers (who can go anywhere and deeply want their work shared) keeps working. But I'm not hopeful. Eventually Llama-n will remain closed source, some researchers will leave, followed by a blog post word salad about safety -- or whatever in vogue pretext works best tomorrow. Ugh.
It's not nearly a comparable situation.
I think we'll continue to see n-1 open models used to build momentum to closed n models for the foreseeable future.
Which is fine, as it will accelerate to diminishing returns on the +1 difference.