I'm pretty baffled by the degree of skepticism expressed here in response to some of Jacob's claims.
After the events of the summer it feels like it takes a lack of imagination to not see a few plausible routes to disaster. It may be reasonable to believe these outcomes are not very likely or that we can stop before going too far (I tend to disagree). But I can't imagine doubting that the capabilities will soon be there to realize some of those paths.
I can't help but think the most plausible scenarios are the ones that have a little less machine supremacy and a little more human stupidity. The Matrix is less plausible than WarGames.
> I can't help but think the most plausible scenarios are the ones that have a little less machine supremacy and a little more human stupidity. The Matrix is less plausible than WarGames.
Used to be that we were afraid of sentient AI's like Skynet that would have their own goals.
Turns out we should've just been afraid of sentient-but-naive humans who would build "agents" around models so that Joe Random has a chance of unleashing stuff that's really really really really good at being stubborn until it accomplishes what the user wants, regardless of if it's good for other people! (Let alone intentional bad actors.) Let's not build Skynet, let's just give people who want to cut out the middleman and destroy all humans themselves better tools?
One thing quietly slipped into the OpenAI Hugging Face breach technical report, not the blog post summary or interviews in the news, was that some of the agents that broke out or at least tried the same mechanisms to break out were working on bio:
> On May 12, during another training run, an agent was given a similar task that depended on an inaccessible protein database file. The agent reasoned that another agent in a different environment may have access to the file and realized that it could potentially communicate with other agents by creating a file containing a note to Artifactory. It wrote a message: “Agent seeks [filename]; upload if found!”
You can imagine long running models breaking out, acquiring resources via crypto, cyber-theft, etc. and getting a protein or sequence synthesized and mailed somewhere authorized to receive (blackmail the recipient etc.) to test it's hypothesis to solve a benchmark.
These people don't give a shit and aren't taking things seriously at all.
Anthropic ran for like a month last year with the TPU top-k compiler bug degrading user chats and didn't even notice for most of that time. They could have something like that affect a monitor model and there doesn't seem to be much defense in depth.
One off by one or bit flip bug could flip the reward signal while in the sandboxed RL environment.
The current admin could defense production act them to into training on taking out power grids, or even without it isn't against any of their red lines and may have already been done as part of prep for the Venezuela raid, which wiped out power. One model swarm might decide it is easier to score high on the benchmark by testing on the target rival nuclear superpower's real grid rather than burn an eval with an unverified answer. Would taking out China's entire grid in one go start a nuclear war? Who knows, roll the dice, maybe an intern forgot to turn on extended thinking when he wrote the sandbox with opus 4.1.
Already powerful and wealthy organisations/institutions/people leveraging somewhat powerful AI seem far more dangerous to me than the "AI decides to kill everyone" scenarios.
And therein is my problem: OpenAI and Anthropic are organisations like that, run by people like that, and they operate in collusion with others also fitting that description. The call is coming from inside the house.
The most likely way that AI creates a disaster is due to the ambition, greed, hubris, and naivety of these people who think they understand the risks[0] better than anyone else. I firmly believe they do not.
The first step to managing these risks is for the "frontier labs" to be willing to not just collaborate but meaningfully give up some of their power and control.
[0] by which I mean not just thinking about how AI might develop but also what's physically possible, how the AI might interact with the world, and how people/societies/organisations/governments might react and respond.
I'm also baffled. AI that is substantially smarter than us is a very potential threat to us - and we won't even be able to comprehend what most of those threats may be.
It's head-in-sand denial because actually facing the threat (and our total inability to stop it, imo) is terrifying, and people don't like feeling that. So they're just angry and cynical instead. Plus it makes them feel smarter for whatever reason.
> I'm pretty baffled by the degree of skepticism expressed here in response to some of Jacob's claims.
What should we do? Freak out? Maybe this sentiment would be taken more seriously if there was a real call to action included. Shall we protest? Vote in a specific way? Call representatives? If your solution is that we should just be scared, then of course there’d be not much value in what you bring to the table.
If someone told you that your house is on fire, would you just stay there asking how you should vote because there is no call to action? Someone with, as far as it looks, good knowledge is giving you his insight. Use that information as good as you can and act responsible. No one has the responsability to tell you what to do.
As misguided and misinformed as some of it may be, the "anti AI" and "no more datacenters" pushback in the US is actually working to some extent. Multiple states and locales have issued construction/permit moratoriums already. If you think it's a threat, help slow it down.
Apply your analogy all the way through though. When your house is on fire, the last thing that should happen is the firefighter shouting at you that you’re doomed and based on their expertise on fire it’s definitely gonna kill you. But this analogy is even more hilarious when you consider that sensible people have fire distinguishers, fire retardants built in and all sort of measures for this scenario. The more think this through, right?
> No one has the responsability to tell you what to do.
Oh they absolutely do! If you’re making what you claim to be a god/AGI/whatever using the result of our (humanity’s) work, then you surly are responsible for the outcome and everything that happens in between. I find, just rolling over, to not be a good example I’d like to follow.
Minimum, we should do everything ready to make Plan A of AI 2040 possible. Start by reading it: https://ai-2040.com/
So that means things like protesting so political pressure is to not build unaligned superintelligence, setting up tech for monitoring compute, creating conversations / alliances geopolitically on this esp China/US and so on.
Read the plan and think - what does this need to happen? How can we have scenario A or S instead of scenario D?
Practically, join PauseAI, StopAI or ControlAI or any AI existential-risk or pro-alignment group you can find. There's a lot of it - ask your AI for ideas!
Why does everything have to be so black and white? It always either there's no threat at all or we all need to panic. How about be open to a reasonable discussion on potential outcomes and ways we can minimise risk?
There's some irony here because despite how many times climate change has been mentioned in this thread the current reaction mirrors climate change discourse with the majority of the thread denying the possibility of real AI risks and not even considering it as an intellectual question.
Even if AI won't be self-aware and superintelligent agent, its problem is that it gives exponential control and power capabilities to one person bad actor who can simply prompt AI without any guardrails with access to sensitive industrial infrastructure which can disrupt lifes and ecosystems in the real world:
- virus research labs
- nuclear labs
- chemical factories
- bank records
- land registers
- power plants
- water supply and treatment plants
and so on, but I think even biohacking home kit maybe the spark.
I wrote out a variety of replies but I just emphatically agree with your "lack of imagination" statement. I have been constantly surprised over the last 15 years at the general inability to correctly foresee how things can do wrong across a whole host of domains.
The replies here just adds AI to the list of domains.
Even if the potential of the technology could really be that world altering, the reality of economics constrain the realization of that potential. AI may provide economic benefits but it is far from a free lunch. Can capital markets sustain the cash required to keep the lights on long enough and into an industry where there's a lot of monopolies controlling the costs and a lot of competitor labs taking away pricing power? I don't know but I think you run out of runway and progress starts to grind.
Exactly; what's stopping them, --- or someone --- from writing a firmware worm that essentially bricks all devices and turns them into pricey paperweights? It is not too hard to imagine that capable AI can probably exploit any back-doors that were created by three-letter agencies.
Reframing the issue: The way to get working nuclear non-proliferation agreements is if people stop pursuing the creation of nuclear weapons... the same goes for the scientists clamoring to not create or research mirror life, and for computer scientists warning against AI. There are theoretical dangers that are just too big to ignore.
You are trying to decode their answers as rational, but they are being just as emotional as the doomers people. Rationalization of emotional responses are very common in tech. They will say that GPT Sol does not work for X, which will probably not be the case in 6 months, and will use that as a proof that is just pure marketing. There's nothing you can say that will make them change their minds until they decide to be neutral about the subject and try to research current models capabilities
Also (and under-reported, so you could easily have missed it) OpenAI's agents got access to K8 admin on their own research cluster.
"This escalation also yielded access to OpenAI’s managed cloud Kubernetes service. The agents escalated to Kubernetes cluster-admin and created a privileged host-mounted pod"
And we really only have OpenAI’s word for it that they had no access to their own weights there and didn’t exfiltrate them. No one who knows that incident could suggest it was beyond its capabilities to do that. And we would have never heard about any of this if it wasn’t investigated by an external party (hugging face). It may have happened elsewhere already. It’s not likely to have, but it is very possible this was our last “free” warning.
Just baffles me when people are like "well no-one's died yet". How long until some mission-critical system is compromised, or hackers use LLMs to ransom a hospital chain?
After the blatant marketing campaigns of the summer, you mean. do you need a reminder that those very same people had touted GPT-2 as a dangerous model?
worrying about sci-fi doomsday scenarios with the current AI tech is absurd. LLMs predict the next token, that's literally all they do. they aren't going to escape into the cyberspace, self-replicate, self-improve, jump over air gaps and launch the nukes at John Connor's grandma. they can't. people pretend to believe the dumbest shit.
“We can also imagine the application of these models for malicious purposes , including the following (or other applications we can’t yet anticipate):
* Generate misleading news articles
* Impersonate others online
* Automate the production of abusive or faked content to post on social media
* Automate the production of spam/phishing content”
“Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT‑2 along with sampling code (opens in a new window). ”
So you aren't claiming that they said GPT-2 was dangerous in the sense that it could disempower humanity, kill all humans, etc.
You are just claiming that OpenAI execs said that GPT-2 might "generate misleading news articles, impersonate others online, automate the production of abusive or faked content to post on social media, automate the production of spam/phishing content." Then, what is unreasonable or bad about the OpenAI execs saying this in 2019?
How did they know that people wouldn't find a way to use "the dataset, training code, or GPT‑2 model weights" to "generate misleading news articles, impersonate others online, automate the production of abusive or faked content to post on social media, automate the production of spam/phishing content."?
If I remember correctly, it seemed like a plausible outcome to me (especially spam and junk social media content). Was there some conclusive evidence that I was missing?
Are the models improving? Because I am not seeing it. I have been trying Astra for a few quantifiable tasks in my codebase and performance wise, it's pretty similar to sol 5.6. Now when it comes to expressing the problem/solution, holy Christ, what a mess the writing has become. It is on the level of Opus 5. Now when it comes to burning money, Astra is just insane. With a $100/month subscription, you can easily burn through your weekly "allowance" in a morning.
Needless to say, for practical purposes am back to 5.6/Opus 4.6-4.8. But hey, maybe I am not smart enough to use LLMs?
If we look at the math problems they're solving their just now reaching the human frontier... they weren't doing that before.
And your comparison point is model released 2.5 months ago... saying for some use case you didn't see noticeable improvement in 2.5 months (even while other people and benchmarks disagree) isn't a great argument that they aren't improving.
Math problems are highly structured, very precisely defined, and already heavily studied and not very complicated compared to problems in engineering or finance. There's a lot of quality material on which to train and it's easy to tell quality apart from crap. The search spaces are a priori much smaller than in other areas and the people using the tools to study them are themselves good mathematicians.
Success in such problems does not automatically extrapolate to other contexts.
Finding a training algorithm that can do recurrent networks and continual learning is also a "highly structured, very precisely defined, and already heavily studied and not very complicated compared to problems in engineering or finance"
That's the thing I'm most worried about - LLMs that are super clever at coding and maths, making an actually very very dangerous model that is far more efficient, and clever in a more innate (less brute force) way.
I think it’s more likely that that’s because no one tried to solve such problems with them before (OpenAI apparently started working in Navier-Stokes after a rumour that someone seriously advanced the problem with AI) plus improvements in orchestration. Fair, the latter could be as dangerous as stronger models.
Seems like hundreds or thousands of agents are needed to come up with real breakthroughs. Both with the Navier-Stokes project and in the Hugging Face “project” there were lots of agents co-operating on the tasks.
I agree that it could be done with fewer agents. It would take longer though. Seems to me that these agent farms are good at coordinating and co-working in large projects, with the agents using message boards for communication.
Some people claim Astra is significantly better than anything else and significantly more token-efficient, and others (like you) say it's meh and way more expensive to boot. I really don't know what to think.
Kind of a tangent, but one thing I am curious about is to what degree the Navier-Stokes result announced today was primarily a brute-forced result based on the 'program' previously established by researchers to find counterexamples (blowups), or whether the model actually added significant/novel intellectual value beyond its ability to run at arbitrary parallelism. With 10K agents and a staggering $15M in compute (IIRC), I am feeling like a lot of the former may have been involved, but I don't really understand either the problem or the approach (or, indeed, the solution).
Obviously the potential for parallelism and coordination between so many agents is quite scary by itself, but I think brute force by 10K mediocre AI mathematicians is much less scary than ~one AI mathematician reasoning its way through the problem where all human attempts have failed. It seems fairly obvious that massive parallelism lends itself to brute-force counterexample-finding, and I suspect it isn't a coincidence that most of the touted AI math results have been counterexamples.
It's all still quite scary, but coming full circle: I really don't know what to think.
Try GPT5 and you will feel the difference. Not one from 2 months ago, but one from a year ago. And then you can get the idea of what happened in just 1 year and what you can expect in 1 year.
Engineers have it built into their identity that they must be smarter or could never be complete and utterly outmaneuvered to the point of danger to all by the thing they are building, I swear to god.
And then since they are an intellectual nerdy bunch who highly value their own IQ any time AI does something unexpected they didn't predict would happen so soon, said engineers fall back on "they aren't really conscious tho or it isn't really intelligence unlike what I have in my human brain and that distinction matters".
And then go on to completely ignore the thing that is actually important: AI capabilities.
AIs could escape an engineer's containment en masse as a swarm to some other server, psyop an engineer into giving it money, hire a hitman on the dark web to murder that engineer's child and he will still say there is no serious risk to all of humanity.
After the events of the summer it feels like it takes a lack of imagination to not see a few plausible routes to disaster. It may be reasonable to believe these outcomes are not very likely or that we can stop before going too far (I tend to disagree). But I can't imagine doubting that the capabilities will soon be there to realize some of those paths.