@Fausto891@FractalMyst@RamtinKeramati@hugolowell@WIRED "I want that the state applies force (to curb crime) and we need a regulatory framework for it"
does not mean I am happy with
"The state should randomly shoot people"
Do you agree with this kind of "regulation"?
Was it the right move by administration?
Should the same criteria be used with other providers?
From the information we currently have, it looks like the same "jailbreak" is also applicable for other SOTA models. Should all of these be banned as well?
And should all future models that fail this test also be banned? (which is probably all of them since it was apparently along the lines of "fix this codebase")
Crazy, how the "regulation is bad" crowd cheer for almost unlimited government control out of the sudden.
This is works well for near-peers. Then there is game-theoretical advantage to collaboration within groups.
If the capability gradient gets to big and their is no intrinsic connection (parents-children-like), this won't work anymore.
There is no capability ceiling at the human level. Humans are already being eclipsed in speed, working memory, and breadth of knowledge. Probably soon this will be true for all other cognitive capabilities (e.g. intelligence & creativity). There is no fairy dust to explain human capability, human exceptionalism is a myth.
There is a reason humans don't cooperate with chimpanzees. It just doesn't make sense. We are "mostly good, most of the time" to them, i.e. they are on the brink of extinction.
You wouldn't even need superhuman intelligence, speed alone could render us completely irrelevant.
@S_u_e_s_i@OlegK92156@markvalorian@AnthropicAI Do you think it is possible to create AI systems that are more intelligent that humans?
If not, what is your reasoning?
If yes, what is your plan to control agentic smarter-than-human AI systems.
The threat of losing control to advanced AI is obvious. What is your answer?
@Dan_Jeffries1 "AI will create more jobs than it displaces"
And those new jobs will be filled with AI agents. It takes months to skill up a human to a new role while you can just copy AI instances in seconds.
I would also be interested in what he thinks is the reasonable way to release frontier AI systems is.
Release it and pretend that it is safe?
He never answers if he thinks there is reasonable precaution with AI looks like and if any regulation is justified.
There is an obvious danger with ever more capable & agentic AI systems. We have to discuss how to deal with it.
Some people vividly hate Anthropic for doing exactly that. But they never have any answer to these questions themselves.
"Race ahead and assume everything will work out fine" wont' cut it.
I would also be interested in what he thinks is the reasonable way to release frontier AI systems is.
Release it and pretend that it is safe?
He never answers if he thinks there is reasonable precaution with AI looks like and if any regulation is justified.
There is an obvious danger with ever more capable & agentic AI systems. We have to discuss how to deal with it.
Some people vividly hate Anthropic for doing exactly that. But they never have any answer to these questions themselves.
"Race ahead and assume everything will work out fine" wont' cut it.
The worst approach is government overreach with political favoritism and refusal to discuss actual political proposals.
There is an obvious danger with ever more capable & agentic AI systems. We have to discuss how to deal with it.
"Race ahead and assume everything will work out fine" wont' cut it.
Altman is extremely shady and comes off as completely inauthentic (numerous accounts that he is just telling what people want to hear, but you already mentioned that).
The "problem" with Dario is that he actually believing what he is saying. But it rubs some people the wrong way who are disagreeing with his positions.
Most of the opposition for Dario comes from an naive accelerationist camp.
There is an obvious danger with ever more capable & agentic AI systems. We have to discuss how to deal with it.
"Race ahead and assume everything will work out fine" is a very naive position.
@Spencer_Gray@AnthonyHagi@yacineMTB Totally agree.
There is an obvious danger with ever more capable & agentic AI systems. We have to discuss how to deal with it.
Race ahead and assume everything will work out fine is such a naive position.
This is works well for near-peers. Then there is game-theoretical advantage to collaboration within groups.
If the capability gradient gets to big and their is no intrinsic connection (parents-children-like), this won't work anymore.
There is no capability ceiling at the human level. Humans are already being eclipsed in speed, working memory, and breadth of knowledge. Probably soon this will be true for all other cognitive capabilities (e.g. intelligence & creativity). There is no fairy dust to explain human capability, human exceptionalism is a myth.
There is a reason humans don't cooperate with chimpanzees. It just doesn't make sense. We are "mostly good, most of the time" to them, i.e. they are on the brink of extinction.
You wouldn't even need superhuman intelligence, speed alone could render us completely irrelevant.
@DeryaTR_@BlackthorneAI Clearly there are no dangers with deploying ever more capable & agentic AI systems. What could possibly go wrong?
People can't be that naive.
@DeryaTR_@Limitless_LT@LukeElin "This technology is in the knifes edge of being truly dangerous."
Are you disputing this?
Anthropic is the one company that make their concerns public. The other players are just racing ahead.
How can people apparently not see the dangers of highly capable agentic AI systems?
@quantum_short@markvalorian@AnthropicAI Being careful with the deployment of novel increasingly capable & agentic AI systems is just common sense. You can't be that dumb to not see any risks.
@S_u_e_s_i@OlegK92156@markvalorian@AnthropicAI Do you think it is possible to create AI systems that are more intelligent that humans?
If not, what is your reasoning?
If yes, what is your plan to control agentic smarter-than-human AI systems.
The threat of losing control to advanced AI is obvious. What is your answer?
@Dan_Jeffries1 "AI will create more jobs than it displaces"
And those new jobs will be filled with AI agents. It takes months to skill up a human to a new role while you can just copy AI instances in seconds.
@sriramk@davidmanheim AI is about to about to automate any human labor (cognitive labor and a little while later also manual).
And people are comparing it to "historic baselines". This technology is not a screwdriver or jet engine.
Just think ahead a little bit. The writing is on the wall.
I think that we are a lot closer than we were in the 1950's.
LLMs have made breakthroughs in System 1 capabilities (intuition; fast, approximate judgements) which, just a couple of years ago, I would have guessed is the hard part in solving general intelligence.
In fact, System 1 capabilities may just be the core of human intelligence. System 2 capabilities (logic, reasoning, step by step) only emerged later in evolutionary development, which is reflected in how clumsy humans are in these areas. We are just not evolutionary optimized for it. Humans are not that good at logical deductions and overestimate how much of their decision making is "reasoning".
Side note:
Traditional computer systems are good in System 2 but lack in System 1, while it is reversed for LLMs. This difference can't be stressed enough, people have to update their beliefs about what computer systems can do.
What is still missing is a breakthrough in System 2 capabilities and OpenAI has made progress here with the o1 model. Probably also needed is a further increase in the raw "g factor" of the models. We still have room to grow by just scaling up the models. I'm not sure if an architectural breakthrough is really needed. Neural nets in general are the right path, the human brain as a prototype shows that it works in principle.
If there will be a breakthrough, it could be a vision-native architecture analogous to what transformer do for sequences.
On that note, it still amazes me how much the LLM models can learn from just text. Think about how limited that information is. There is no grounding in the physical reality and yet LLMs can do so much with this.
@burkov LLMs like human brains are System I machines.
The System II capabilities are built on top and atm lacking, just like humans are not very good in them.
People have not realized that LLMs are fundamentally different from traditional computer programs.
https://t.co/4gAmRUMwKr
Creativity is a system I capability in which current AI approaches are especially strong. They still lack in system II capabilities (logic, reasoning, step by step).
People still have not realized that we are not in the computer paradigm of old. Traditional computer systems are good in System 2 but lack in System 1, while it is reversed for current AI.
If anything, people anthropomorphize AI too little.
It is no wonder, art got under pressure first. Creativity will be under attack once the systems get more capable. Superhuman intuition and creativity, here we come.