I don’t think they’re blocked by an inability to run autonomously. They’re blocked by lacking an eye for novelty/interestingness. You can make the slop factory to run 24⁄7 for a year and still not get any closer to solving alignment.
IC Rainbow
Karma: 73
says little about the intelligence of Claude
It says that it lacks intelligence to play zero shot and someone has to compensate the intelligence deficit with an exocortex.
It’s like we can track progress by measuring “performance per exocortex complexity” where the complexity drops from “here’s a bunch of buttons to press in sequence to win” to “”.
AIs (probably scaffolded LLMs or similar)
That was a good start, but then you appear to hyper-focus on the “LLM” part of a “blogging system”. In a strict sense the titular question is like asking “when will cerebellums become human-level athletes?”.
Likewise, one could arguably frame this as a problem about insufficient “agency,”
Indeed. In a way, the real question here is “how can we orchestrate a bunch of LLMs and other stuff to have enough executive function?”.
And, perhaps, whether it is at all possible to reduce other functions to language processing with extra steps.but it is mysterious to me where the needed “agency” is supposed to come from
Bruh, from the Agancé region of France of course, otherwise it’s a sparkling
whileloop.
Why fire devs that are 10x productive now and you can ship 10x more/faster? You don’t want to overtake your unaugmented competitors and survive those who didn’t fire theirs?
I wondered about using 4o for the poll and took the post to o1-pro.
Here’s what it filled as “Potential Gaps or Additions”:
Deceptive Alignment / Treacherous Turns
The final report mentions partial or “if-else” compromises (where the AI does X in some contexts and Y in others), but they don’t explicitly name the risk that the AI behaves aligned until it’s powerful enough to defect.
That might be implied, but it isn’t spelled out as a well-known scenario. Some alignment taxonomies place “deceptive alignment” in its own major category, since it’s so crucial.
Multi-Agent or Subculture Effects
The scenario does mention “200,000 copies of the agent working in parallel,” but the final part doesn’t dwell much on possible emergent subcultures or stable “group norms” that could shape the AI’s goals. Hypothesis 6 (Other) could cover it, but it’s not elaborated.
In large multi-agent systems, it’s possible for goals to drift or unify in unexpected ways—like AIs forming an in-group identity or hidden coordination.
Lack of a Standalone “Self-Preservation / Power-Seeking” Hypothesis
The final part lumps instrumentally convergent drives (power, self-preservation) under Hypothesis 5 as a special case of “proxies.” That’s not incorrect, but some alignment researchers call out “power-seeking” as a separate major outcome, given how often it arises in discussions about advanced RL.
Absence of “No Stable Goal” Scenario
They do not explicitly mention the possibility that the system remains relatively non-goal-driven or “policy-like.” Some argue that a sufficiently large RL agent might not converge to a monolithic goal at all, but rather remain a complex patchwork of heuristics with no coherent objective.
That said, Hypothesis 3’s partial alignment or Hypothesis 6’s “Other” might implicitly allow for a “messy mixture” that never coalesces into a stable final goal.
Explicit Mention of “Identity Formation”
The final part briefly references the idea that large language models can pick up “role-playing” from pretraining data, but it doesn’t dwell on the phenomenon where an AI might choose or synthesize a stable persona that is not exactly any of the training signals. This is arguably a sub-point of Hypothesis 6.
The full history has two tables for credences (the main dish and the extras) with unclear provenance. To spice things up I’ve also asked for expected evidence to update up/down.
human-made innovative applications of the paradigm of automated continuous program search. Not AI models autonomously producing innovations.
Can we… you know, make an innovative application of the paradigm of automated continuous program search to find AI models that would autonomously produce innovations?
RL will be good enough to turn LLMs into reliable tools for some fixed environments/tasks. They will reliably fall flat on their faces if moved outside those environments/tasks.
They don’t have to “move outside those tasks” if they can be JIT-trained for cheap. It is the outer system that requests and produces them is general (or, one might say, “specialized in adaptation”).
Reality, unlike fiction, doesn’t need to have verisimilitude. They are persuaded already and racing towards the takeover.
What’s the last model you did check with, o1-pro?
For alphazero, I want to point out that it was announced 6 years ago (infinity by AI scale), and from my understanding we still don’t have a 1000x faster version, despite much interest in one.
I don’t know the details, but whatever the NN thing (derived from Lc0, a clone of AlphaZero) inside current Stockfish is can play on a laptop GPU.
And even if AlphaZero derivatives didn’t gain 3OOMs by themselves it doesn’t update me much that that’s something particularly hard. Google itself has no interest at improving it further and just moved on to MuZero, to AlphaFold etc.
The first one, why?
Do you have a more concrete example? Preferably the one from the actual EA causes.
How one should signal their decision procedure in real life without getting their ass busted for “gambling with lives” etc.?
Getting stuff formally specified is insanely difficult, thus unpractical, thus pervasive verified software is impossible without some superhuman help. Here we go again.
Even going from “one simple spec” to “two simple spec” is a huge complexity jump: https://www.hillelwayne.com/post/spec-composition/
And real-world software has a huge state envelope.
Even if that’s the case, the amount of 0-days out there (and just generally shitty infosec landscape) is enough to pwn almost any valuable target.
While I’d appreciate some help to screen out the spammers and griefers, this doesn’t make me feel safe existentially.
Eliezer believes humans aligning superintelligent AI to serve human needs is as unsolvable as perpetual motion.
I’m confused. He said many times that alignment is merely hard*, not impossible.
Especially with current constraints from hell.
I’m getting the same conclusions.
Think of a company like Google: building the biggest and best model is immensely valuable in a global, winner-takes-all market like search.
And this is in a world, where Google already announced that they’re going to build even bigger model of their own
We are not, and won’t for some* time.
We have to upgrade our cluster with a fresh batch of Nvidia gadgets.
I doubt that any language less represented than English (or JS/Python) would be better since the amount of good data to ingest would be much less for them.
When we evo-pressure visibly negative traits from the progressively capable AIs using RLHF (or honeypots, or whatever, it doesn’t matter), we are also training it for better evasion. And what we can’t see and root out will remain in the traits pool. With time it would be progressively harder to spot deceit and the capabilities for it would accumulate at an increasing rate.
And then there’s another problem to it. Deceit may be linked to actually useful (for alignment and in general) traits and since those would be gimped too, the less capable models would be discarded and deceitful models would have another chance.
presumably while starting out friendly
I don’t think it can start friendly (that would be getting alignment on a silver plate). I expect it tostart chaotic neutral and then get warped by optimization process (with the caveats described above).
Confused how?.. The only thing that comes to mind is that’s FOOM sans F. Asking for 0.2 FOOMS limit seens reasonable given current trajectory 😅
That’s because they can read its thoughts like an open book.