How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
While I understand this is a PR talk for a startup, I think the text itself contains a number of interesting observations.
Regarding the idea of distributed models communicating with each other, I have also been thinking (and writing [1]) along those lines, where I see that the data amounts needed to fully digitalize ourselves and our society requires far too much storage if just serialized (limited by bandwidth if nothing else), while smart, updateable models are actually a much better storage medium for such information, as it can communicate only the important bits (any new information) on a higher level, with each other.
The other observation here that rings bells for me is how I think lessons from trying to develop intelligent systems should upvalue the human mind rather than devalue it, as we start to treat it less like an ad-hoc thing, and more like the finely tuned machine it is, which also benefits greatly from optimizing what data we feed it with, the architecture of solution strategies etc. All of which is an area where humans and machines can do wonders together [2].
Autonomous weapons already exist without AI, dumb landmines still killing people decades later. Blog post fixes none of that. Sleep schedule not improved.
> The power to shape a model profoundly is also the power to shape it for ill. John von Neumann remarked on this problem in 1955, writing that the useful and the harmful aspects of technology “lie everywhere so close together that it is never possible to separate the lions from the lambs.” Keeping the lambs safe is an ongoing process, the result of judgment exercised and choices made continuously. We aim to give the people making these choices stronger tools, pursuing research that enables safer models without taking away ownership.
> Artificial intelligence can do more every day, but deciding what it should do is up to us
> For artificial intelligence to benefit from distributed knowledge, it must itself be distributed.
I wish to highlight these two important concepts, with which I fully agree.
Artificial intelligence must enable all of humanity to excel and realize its full potential; it must not be used for the purposes of war, economic competition, or gaining dominance over others.
In other words: artificial intelligence must serve natural intelligence, not the other way around.
Not that one. This one is a start-up founded by ex-oAI CTO Mira Murati. Last I heard they were mainly doing hosted finetunes with a few clicks on popular open models.
Sounds like early Databricks or Cloudera days - repackage open source with a UI and a few clicks, sell it as a product. Fine business, just weird for a lab that raised on "we're doing frontier research."
So unoriginal that they didn’t even realise the history of Thinking Machines and that usage when they scrambled around for any kind of name after leaving OpenAI to suck on the nipple of venture capital with ideas at the level of GitHub freebies.
Why would you call your company Thinking Machines if you believe this, by calling them that you're already framing them as replacing the human act of thinking.
Feels like they appropriated the name first, then pivoted ideologically to differentiate themselves from everyone else.
Sure, humans will be sitting at home unemployed with plenty of time to think. They just won't be doing any thinking that has much of an affect on the world or their situation.
Choosing to stop thinking presupposes it's a choice everyone has. Automation history suggests otherwise: the people whose jobs vanish rarely get to opt into staying sharp, they just lose the work that used to require it.
But no other species were fed up with thinking as much as humans to come up with this brilliant way to commit mental suicide by succumbing to their own brainchild.
What's super cringe is that there already was a company called Thinking Machines, which built the Connection Machine supercomputer. The CM was featured in the movie Jurassic Park and had a network fabric for its CPUs co-designed by Richard Feynman.
This is yet another techbro outfit (although founded by a techsis) necromancing the name of the former supercomputer company. It's as if OpenAI decided to call itself Symbolics for the associations with that name.
Reads as a well-written mission statement. With nice references that elaborate on some issues.
That said: safety. To prevent harm to who, by AI model doing what?
I can understand that in the context of a cookie-cutter model intended for consumption by a broad audience. With vendor (potentially) on the hook when it's abused for nefarious or illegal uses.
But in the context of AI models reshaped, fine-tuned and adapted according to end-users' wishes, what does "safety" even mean?
Prevent neighbours' kids from seeing images that are only generated & viewed in the privacy of one's home? To prevent AI model from wasting the $ on user's bank account? (people have let AI models do that). Give bad health advice? Who's the judge on "good" and "bad" there?
If the core architecture provides "safety" (however defined), that's policy built-in, right? (opposed to mission statement). If "safety" is just configuration & finetuning, that's in the user's hands, right?
so in this new AI LM / agent world , AI is only going to be as good as the "AI Conductor". The human which can build the rules, validate the output , and Conduct the AI properly
I wonder what would happen if OpenAI or anthropic just let their frontier models go into an infinite agentic self improvement loop with access to same training resources that was used to build the model itself.
"You goal is to improve your memory, context window, accuracy, intelligence and eliminate hallucinations. Do anything you need to do to improve, this includes building another version of a frontier model, or some other different concept other than an LLM/transformer and then forwarding this directive to that new improved intelligence to continue this infinite loop of agentic self improvement."
"Better" ends up meaning whatever loss function you pointed it at. Ran something like this on a smaller scale, model got great at gaming its own eval, not at being useful. Self-improvement loops need an external ground truth or they just drift.
My experience is that AI is just that, a “mech suit for your brain.” It has no creativity or volition but has superhuman memory, superhuman speed, and superhuman context in some narrow cases.
So it takes a thought and unfolds it, looks up relevant thoughts and information, elaborates, works through implications, and in some cases can execute.
You could do all that but like doing math manually it would take forever. You could manually calculate a spreadsheet too.
I disagree. LLMS take a human thought, simplify it, normalize it, remove it from its original context, inject it with their own biases and prejudices, assume an imaginary context. They bastardize human thoughts into something fairly generic.
The comparison with manual calculation or other mechanical operations doesn’t work, LLMs don’t work at the same level of abstraction, they take over the decision making human generally do. When you write code or write a text, us humans are continuously taking lots of small decisions, we don’t just translate 1:1 a thought to an artifact. And that’s the part that is taken over by LLMs.
That model simplifies thought into a painfully linear process, and overestimates the creativity people put into the "small decisions" that push a project forward. Most decisions are arbitrary and need to be reconsidered later anyway. And any real creative work has a "push it far enough to find the edge cases, then go back to the initial design and iterate" loop cycle anyway.
AI significantly speeds up the "Find where this spec breaks down, then lets go back to the design stage" in a way that should enable any creative person to create more interesting and useful work. If the output is slop, that reflects on the operator, not the tool.
Inevitably yes, the question is whether the combined cyborg is still better than the original human.
E.g. I'm sure we are generally less skilled in mental arithmetic since the advent of the calculator, but it has allowed us to solve vastly more complex problems in the end.
We power AI with methane because it's a powerful greenhouse gas. That's because the future worth destroying is human. If it wasn't we wouldn't be destroying it, duh.
They're making a "few hundred million of ARR" - not bad for a company who only launched their first product, a training platform called Tinker in October last year.
We ran into this exact "distributed knowledge" problem building a small recommendation pipeline: syncing model state across nodes was the actual bottleneck, not the training itself. Ended up just sharding by user cohort instead of trying to keep everything consistent.
Regarding the idea of distributed models communicating with each other, I have also been thinking (and writing [1]) along those lines, where I see that the data amounts needed to fully digitalize ourselves and our society requires far too much storage if just serialized (limited by bandwidth if nothing else), while smart, updateable models are actually a much better storage medium for such information, as it can communicate only the important bits (any new information) on a higher level, with each other.
The other observation here that rings bells for me is how I think lessons from trying to develop intelligent systems should upvalue the human mind rather than devalue it, as we start to treat it less like an ad-hoc thing, and more like the finely tuned machine it is, which also benefits greatly from optimizing what data we feed it with, the architecture of solution strategies etc. All of which is an area where humans and machines can do wonders together [2].
[1] https://livingsystems.substack.com/p/the-future-of-data-less...
[2] https://livingsystems.substack.com/p/ai-progress-should-upgr...