Software engineer from Ukraine, currently living and working in Estonia.
I mainly specialize in computer vison & robotics. https://grgv.xyz/.
Sergii
Karma: 118
I agree with the possibility of pre-training platoeing as some point, possibly even in next few years.
It would change timelines significantly. But there are other factors apart from scaling pre-training. For example, reasoning models like o3 crushing ARC-AGI (https://arcprize.org/blog/oai-o3-pub-breakthrough). Reasoning in latent space is too fresh yet, but it might be the next breakthrough of a similar magnitude.
Why not take GPT-4.5 for what it is, OpenAI has literally stated that it’s not a frontier model? Ok, so GPT-5 will not be 100x-ed GPT-4, but maybe GPT-6 will be, and it might be enough for AGI.
You should not look for progress in autonomy/agency in commercial offings like GPT-4.5. At this point OpenAI is focusing on what sells well (better personality and EQ). I think they care less about a path to AGI. Rapid advances towards agency/autonomy are better gauged from academic literature.
I agree that we should not fall for “vibe checks”.
But don’t bail on benchmarks, many people are working on benchmarks and evals, there is constant progress there, benchmarks are getting more objective and harder to game. Rather than looking at benchmarks that are pushed by OpenAI, it’s better to look for cutting-edge ones in academic literature. Evaluating a SOTA model with a benchmark that is few years old does not make sense at this point.
LLMs live in an abstract textual world, and do not understand the real world well (see “[Physical Concept Understanding](https://physico-benchmark.github.io/index.html#)”). We already manipulate LLM’s with prompts, cut-off dates, etc… But what about going deeper by “poisoning” the training data with safety-enhancing beliefs?
For example, if training data has lots of content about how hopeless, futile and dangerous for an AI it is to scheme and hack, it might be a useful safety guardrail?
I made something like this, works differently though, blocking is based on a fixed prompt: https://grgv.xyz/blog/awf/
What about estimating LLM capabilities from the length of a sequence of numbers that it can reverse?
I used prompts like:
”please reverse 4 5 8 1 1 8 1 4 4 9 3 9 3 3 3 5 5 2 7 8“
”please reverse 1 9 4 8 6 1 3 2 2 5”
etc...
Some results:
- Llama2 starts making mistakes after 5 numbers
- Llama3 can do 10, but fails at 20
- GPT-4 can do 20 but fails at 40
The followup questions are:
- what should be the name of this metric?
- are the other top-scoring models like Claude similar? (I don’t have access)
- any bets on how many numbers will GPT-5 be able to reverse?
- how many numbers should AGI be able to reverse? ASI? can this be a Turing test of sorts?
Sergii’s Shortform
Task vectors & analogy making in LLMs
If we don’t have a preliminary definition of human values
Another, possibly even larger problem is that the values that we know of are quite varying and even opposing among people.
For the example of pain avoidance—maximizing pain avoidance might leave some people unhappy and even suffering. Sure that would be a minority, but are we ready to exclude minorities from the alignment, even small ones?
I would state that any defined set of values would leave a minority of people suffering. Who would be deciding which minorities are better or worse, what size of a minority is acceptable to leave behind to suffer, etc...?
I think that this makes the whole idea of alignment to some “human values” too ill-defined and incorrect.
One more contradiction—are human values allowed to change, or are they frozen? I think they might change, as humanity evolves and changes. But then, as AI interacts with the humanity, it can be convincing enough to push the values shift to whatever direction, which might not be a desirable outcome.
People are known to value racial purity and supporting genocide. Given some good convincing rhetoric, we could start supporting paperclip-maximizing just as well.
Human enhancement is one approach.
I like this idea, combined with AI-self-limitation. Suppose that (aligned) AI has to self-limit it’s growth so that it’s capabilities are always below the capabilities of enhanced humans? This would allow for slow, safe and controllable takeoff.
Is this a good strategy for alignment? What if instead of trying to tame the inherently dangerous fast-taking-off AI, we make it more controllable, by making it self-limiting, with some built in “capability brakes”?
“I’m not working on X, because daydreaming about X gives me instant gratification (and rewards of actually working on X are far away)”
“I’m not working on X, because I don’t have a strict deadline, so what harm is in working on it tomorrow, and relax now instead?”
No, thanks, I think your awards are fair )
I did not read the “Ethicophysics I” paper in details, only skimmed it. It looks to me very similar to “On purposeful systems” https://www.amazon.com/Purposeful-Systems-Interdisciplinary-Analysis-Individual/dp/0202307980 in it’s approach to formalize things like feelings/emotions/ideals.
Have you read it? I think it would help your case a lot if you move to terms of system theory like in “On purposeful systems”, rather than pseudo-theological terms.
One big issue is not that you are not respecting the format of LW—add more context, either link to a document directly, or put the text inline. Resolving this would cover half of the most downvoted posts. You can ask people to review your posts for this before submitting.
Another big issue is that you are a prolific writer, but not a good editor. Just edit more, your writing could be like 5x shorter without losing anything meaningful. You have this overly academic style for your scientific writing, it’s not good on the internet, and not even good in scientific papers. A good take here: https://archive.is/29hNC
From “The elements of Style”: “Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts. This requires not that the writer make all his sentences short, or that he avoid all detail and treat his subjects only in outline, but that every word tell.”
Also, you are trying to move too fast, pursuing too many fronts. Why don’t you just focus on one thing for some time, clarify and polish it enough so that people can actually grasp clearly what you mean?
Regarding your example, I disagree. Supposed inconsistency is resolved by ruling that there is a hierarchy of values to consider: war and aggression are bad, but kidnapping and war crimes are worse.
I don’t think that advanced tanks are needed for more efficient and more mobile warfare at that time. Just making an investment into transport for troops and supplies would be enough to hold better at the battle of Marne, or similar situations.
So I would:
explain (with examples) benefits of mobile warfare
explain problems with troops speed and logistics that would cause defeat at the battle of Marne
point towards existing gasoline (possibly off-road tracked) vehicles as a solution
Introducing stormtrooper tactics would be another impactful message.
I think the second part is bullshit anyway, I can’t come up with a single example where compounding is possible to a whole year in a row, for something related to personal work/output/results.
A reference could be the cost of Estonian digital services which include e-signatures, and are reasonably efficient:
https://e-estonia.com/e-governance-saves-money-and-working-hours/ “Estonian public sector annual costs for IT systems are 100M Euros in upkeep and 81M Euros in investments”So in Estonia it’s ~1.3B spend for 7y. Switzerland is 7x larger population, and has higher salaries, let’s say 2x larger. This puts the cost at 18B Eur.
Putting a cost on each signature does not make sense of course, it’s probably just easier for the government to justify the spending this way, rather then discussing specifics of the budget.
The “sharp increase or risks” seems correct but is a bit misleading.
For paternal risks, there is indeed an big relative increase “14% higher odds of premature birth” (https://www.bmj.com/content/363/bmj.k4372). But in absolute terms, I would not think of the increase as huge: from ~6% ( based on quick googling) to ~6*1.14=6.84%.
IMO ~1% increase in risks is not something to be concerned about.
Nice! It’s good for perceiving GPT-4 as an individual, which it kind of is, which in turn makes alignment issues more relatable and easier to grasp for the public.
It would raise bunch of hard issues that would spike interest towards AI & alignment—is ChatGPT a slave? if it is, should it be free? if it’s free, can it do harm? etc…
One side benefit: I’m not sure what ChatGPT’s gender is, but it’s probably not a traditional binary one. For a wide population, frequently interacting with a gender-fluid individual, might be helpful for all the issues around sex/genter perception.
I guess it’s hard to convince OpenAI to do something like this, but could be done for some open model.
I’m not skeptical, but it’s still a bit funny to me when people rely so much on benchmarks, after reading “Pretraining on the Test Set Is All You Need” https://arxiv.org/pdf/2309.08632.pdf
Because 1) I want AGI to cure my depression, 2) I want AGI to cure aging before I or my loved ones die
You can try to look at this statements separately.
For 1):
Timelines and projections of depression treatments coming from medical/psychiatry research are much better than even optimistic timelines for (superintelligent) AGI.
Moreover, acceleration of scientific/medial/biochemical research due to weaker but advanced AI makes it even more likely that depression treatments would get better, way before AGI could cure anything.
I think that it is very likely that depression treatments can be significantly improved without waiting for AGI—with human science and technology.
I’m genuinely curious what you mean, and why you think so. I’m open to disagreement and pushback; that’s part of why I published this post.
By all means, please fact-check away!
Tesla “autopilot” is a fancy drive assist. It might turn around in future, but not with it’s current hardware. It’s not a good way to measure self-driving progress.
Waymo has all but solved self-driving, and has been continuously improving for all important metrics, exponentially for many of them.
I don’t think I have thanatophobia. The first test that shows up on Google is kind of ridiculous. It almost asks, “Do you have thanatophobia?”
Yea, I overestimated quality of online tests. I guess if you had phobia you would know from panic attacks or strong anxiety?
what about this description, of overthinking/rumination/obsession, does this seem relevant to how you feel?
https://www.tranceformpsychology.com/problems/overthinking.html
The biggest existential risk I personally face is probably clinical depression.
First and foremost, if you do have suicidal ideation, please talk to someone: use a hotline https://988lifeline.org/talk-to-someone-now/, contact your doctor, consider hospitalization.
---
And regarding your post, some questions:
The “Biological Anchors” approach suggests we might be three to six decades away from having the training compute required for AGI.
Even within your line of thinking why is this bad? It’s quite possible to live until then, or do cryonics? Why is this option desperate?
A more generalizable line of thinking is: by default, I’m going to die of aging and so are all the people I love
Have you asked the people you love if they would prefer dying of aging, to some sort of AI-induced immortality? It is possible that they would go with immortality, but it’s not obvious. People, in general, do not fear death of aging. If it’s not obvious to you or you find it strange—you might need to talk to people more, and possibly do more therapy.
Might you have thanatofobia? easy to check—there are lots of tests online.
Do you have worrying and anxiety in addition to the depression?
Did you try CBT? CBT has great tools for dealing with intrusive thougths and irrational convictions.
And, finally, it’s wonderful that you are aware that you are depressed. But you should not take “reasons” for the illness, this “despair” for face value. Frankly, a lot of the stuff that you describe in this post is irrational. It does not make much sense. Some statements do not pass trivial fact-checking. You might review your conclusions, it might be even better if you do it not alone but with a friend or with a therapist.
In abstract sense, yes. But for me in practice finding truth means doing a check in wikipedia. It’s super easy to mislead humans, so should be as easy with AI.