An OpenAI employee has observed that Large Language Models starting with the same dataset converge to the same point. This would mean curating the data is the critical step in creating safe ASI (Artificial Super Intelligence). If we can front load the AI with the desired ethical and pro-humanity examples, then the resulting AI system should have a core of human aligned safety.
interesting post from an OpenAI employee claiming that all large language models reach the same endpoint regardless of training strategy or clever tricks
this is of course what the bitter lesson teaches us but useful to get an up to date confirmation that it still holds true pic.twitter.com/kpBooKyqm4
— Andi 🇦🇱🇺🇦 (e/alb) (@Nexuist) November 23, 2023
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Premise here is that a scaled-up LLM is all you need for ASI. There are some AI researchers who think so but others think we still need fundamental breakthroughs. If an ASI is not just an LLM then these results won’t apply.
I don’t think you CAN scale up an LLM large enough. The training data available just isn’t massive and diverse enough.
And I think the hallucination problem isn’t a problem of not enough data or training, it’s fundamental: These systems don’t understand anything. They’re just enormously complex auto-complete algorithms, essentially.
A genuine AI needs to have internal models that amount to understanding facts and concepts, not just a statistical model of text, however complex.
I Tweeted:
Sounds like it’s the food, actually. Also, GIGO still applies & AI doesn’t do introspection & self-questioning of beliefs very well if at all. It’s really SI – Simulated Intelligence – not AI.
This is true. The only problem is that the term “Artificial Intelligence” being applied to these new systems, while misleading enough to make people think we are on the verge of developing AGI when we haven’t gotten real AI yet seems almost necessary. It is necessary because by not using the term “AI” it makes it hard to convey how much power and promise this technology has. Just being able to program a robot using video instead of manually coding every action and reaction is going to be transformational and the better the simulations of intelligences become the more we will realize just how much can be accomplished without a fully intelligent entity. We just didn’t know this because until now the only way to do many things was to get a genuinely intelligent entity (or whatever humans count as) to do it.
So we either use a dishonest term and mislead ourselves or we don’t use the term and remain unaware of what is happening, thus perhaps misleading ourselves more so.
I think people WANT to believe that LLM’s are actual AIs, because the alternative is recognizing that most human activities can be adequately copied without much intelligence or genuine understanding being needed at all. Which is kind of embarrassing to realize.
For humans, actual intelligence is an error handling routine, not a normal mode of operation. It pretty much has to be, it’s rather computationally inefficient for normal tasks.
This is assuming that some awful person doesn’t front load the AI with anti-Human ethics.
There are plenty of awful people and drivers to do this.
AND, given some of the bizarre things people in the tech community sometimes think are ethics, that seems almost inevitable.
Tony notes the recent hubbub about ChatGPT refusing to save a billion white people if it took uttering a racial slur. I doubt the programmer who commanded the program could never utter racial slurs really mean “even if it would save a billion lives”, but they gladly programmed this outcome in anyway.
Human inconsistency often saves us from the inhuman implications of stupid ethical beliefs. Not every self-proclaimed “Marxist” ends up trying to create a secret police and gulag. But an AI with Marxism in the training data would deterministically end up there every time…
indeed there are bad people & ideas in the industry.
If we can’t ban moral relativism (in which the ends justify the means!!!) & it’s subsidiaries from AI ethics programming, then we must program all good AI (& not just a few) to automatically “combat” the inevitable problems that will spring up from such unethical programing.
I asked chatgpt what was worse killing someone or a racial slur and it couldn’t decide what was worse basically saying both where equally bad in their own way. I am worried that woke chatgpt is really problematic as it is morally messed up.