▲ 152 points
back
89 comments
Absolutely unacceptable that their TOS basically gives them unrestricted access to your datasets, as far as I read it. The terms let them use your datasets for pretty much whatever they decide they want to do with it (though they do say they would anonymize the data, which isn't especially helpful). The TOS leaves a lot of wiggle room for them to do pretty much whatever they want to the data.
I wouldn't touch this until they get serious about having real assurances that they're not going to access customer data without a real, justifiable reason. If Amazon gave themselves free reign to read S3 data it would be outrageous, this is basically the same thing.
But hey, it “empowers researchers and hackers to experiment with models by giving them control over the algorithms and data”.
PR and marketing will literally write anything they can get away with. Thanks for pointing out the TOS hole.
>Absolutely unacceptable
Entitled much!? hold on to you pants.
They've clearly called out that organizations can contact them for their specific needs..
Your datasets are some of your most valuable assets. How is it entitled to not want your vendor having a free for all with them? It would be entitled if they weren't going to be charging for it.
That should not be the default
My research group at Stanford has been alpha testing Tinker, it's both very useful and also really technically impressive in my opinion. It's a unified framework for post-training models and it abstracts almost all of the complexity of managing these jobs across resources. That it manages to do this while also allowing a lot of algorithmic flexibility is pretty unique.
Silly question: how is it different from, say, hf's transformers and similar libraries and APIs?
with hf transformers, you still need to manage GPUs
Interesting that their first product is an infrastructure play. Is it really so hard to set up a fine-tuning pipeline for yourself that a $12 billion startup with unlimited hype needs to be offering it? Maybe they have figured, whether correctly or not, that building AI tooling is going to be more lucrative than the AI itself.
The thing about this that’s interesting to me is that it can be used as a foundation for products they or other people make that combine real time RL rewards and fine tuning to improve the model. I see a lot of potential here compared to the standard paradigm of ChatGPT wrappers that involve tweaking the prompt or harness to improve it, which is a lot more constrained.
OpenAI has had a fine-tuning API since GPT-3.5, and a reinforcement fine-tuning API since last year.
"the Tinker API provides simple functions to compute gradients, update the weights, and sample outputs from the trained model"
It sure sounds like a PyTorch tutorial, but I believe it's yet another "AI training made slightly easier for you" start-up. But all of them seem to solve the easy problem of managing data and compute, while very few tackle the hard problem of generating good training data.
Commonly attempting the "private beta with a waitlist" pseudo-release model (until they finally learned their lesson relatively recently) is a large part of how google fumbled the LLM ball to OpenAI and others.
This is great as we need more solutions that help people train and fine-tune models. There is a lot of open source and some companies behind some of the popular open source packages, like Unsloth, but the more the merrier, esp given that Thinking Machines has the expertise and resources to build something that last.
Yeah, I'm really curious about their stacked multi-tenant lora training at the same time. If this gets commoditised enough, it could be interesting to try "end of the day fine-tunes on daily conversations" and see where that leads. Or a targeted RL on "missed / rejected tasks" for an agent, after you get enough samples for a run, and so on.
1. Can someone help me articulate what Tinker can do that Vertex AI or many others can't? (I can see access to some primitives, which is nice)
2. and more broadly: Has anyone got real lift in business metrics through fine-tuning an open model over using the flagship models from say OpenAI or Anthropic?
Most managed finetuning offerings take your dataset, some hyperparameters, and spit out a model. Few support RL, and those that do have very limited support.
And I have gotten a real lift, in cost effectiveness and engagement (for creative writing)
What are you doing that requires RL-ing creative writing for engagement?
I don't apply RL directly to engagement (and don't think it's really possible without some insane scale of feedback)
Instead there are mechanical mistakes models make that harm engagement and are trivially verifiable (overused phrases and concepts, hitting a given target reading level, etc.)
Improving those is what improves engagement.
> Tinker is a flexible API for efficiently fine-tuning open source models with LoRA.
It would be great if they offered inference from the trained model as well. Ideally pay per token.
Given that Thinking Machines has employed so many smart scientists, focusing solely on infra and fine-tuning is kind of a letdown.
I feel like this is what the current team excelled at at OpenAI, only makes sense that they would productize it
Funny timing that they are announcing their first product days after Matt Levine highlighted their lack of a public product or direction in the Money Stuff newsletter.
It's the first that came to mind when I saw the domain name but I'm sure he's not the first to point it out. So it probably is just inevitable that they'd be launching after someone has mentioned it during the last couple of hours/days.
Their lack of a public product has been widely discussed for at least a month now
There was a famous tech company in supercomputing and AI, called "Thinking Machines", worked at by people such as Danny Hillis and even Richard Feynman.
Does this new company have some connection to that, such as some of the same people?
> Does this new company have some connection to that, such as some of the same people?
Really doubt it. This "Thinking Machines" seems like an on-trend SV AI software startup, that "Thinking Machines" was a Massachusetts-based hardware company that went bankrupt 30 years ago (and whatever's left is Oracle). That "Thinking Machines" people are retiring by now.
I wonder if Sun/Oracle let the trademark lapse, because if it's still active I'd imagine this startup's gonna get sued.
No relation AFAIK, this is Mira Murati's post-OpenAI joint.
Which makes the hero image on this page being a diagram of Danny Hillis' TinkerToy computer all the more baffling.
Probably just an homage - tipping the hat to the predecessor
nope, this is Mira Murati's thinking machines. she used to be the CTO of OpenAI, but started her own thing once she [I think] realized that she was missing out on the gold rush by staying there.
They mention their "cookbooks" but I couldn't find them... Their blog was immensely interesting so I could see this being a good entrypoint
This company was most recently valued at $12 billion. Tell me there's not an AI bubble.
https://www.reuters.com/technology/mira-muratis-ai-startup-t...
That's a fair valuation if they have a >10% chance of developing a frontier model. It's an excessive valuation if they just release tools to create LoRAs and such.
They seem to position themselves to commoditise renting shovels to use with your own data. Seems pretty bubble-safe to me.
Isn't this a feature offered by many LLM providers? What's your USP here?
It is? Which ones give you distributed post training of MoE LLMs?
Usually we want an announcement to come with more than a waitlist before doing a frontpage thread on HN, but I guess this company is high-profile enough that the post is relevant anyway?
A/B testing the python scripts
The name chosen is an antiquated ethnic slur in much of the Anglosphere.
I’ve never heard of this in my life. Isn’t tinker a verb meaning to fiddle or edit or modify in small increments? That seems like the perfect name given what the software they’re present. In any case I guess this is back to the debate: does it matter how a word is intended or does it matter how a word is received?
Good thing it's been forgotten enough that almost nobody knows that, or cares it's the name of this company.
I think if anyone is offended by a word that is not used by anyone in that context, they're probably due for some self-reflection on what offends their sensibilities.
It is. But honestly, fuck that. It has a colloquial meaning that carries none of that baggage. Instead of retiring the word, why not make the racist definition archaic?
How much of the Anglosphere?