site banner

Culture War Roundup for the week of December 4, 2023

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

5
Jump in the discussion.

No email address required.

This result shouldn't be underestimated because Gemini-Ultra is merely on par/slightly better in text-based reasoning: it thoroughly beats GPT-4V on MMMU, the multimodal benchmark, including harder subscales; it also plays well with audio. People are for the most part functionally illiterate, so this is huge; and of course they will capitalize on Android and other ecosystem-wide advantages the Alphabet empire has. Multimodal language-like models will obviously be table stakes in 2024. (Bytedance guy even hints that they'll opensource a model on Gemini's level.)

Interesting that one of people who had worked on aligning early Gemini said they had trouble aligning it – it burned through RLHF reward models, finding exploits and collapsing into gibberish (imagine using actual RLHF in 2023!). Maybe this has delayed the release, as well as the garden variety safetyism it has made more complex.

To be honest I was more excited about the other day's release of Mamba by Albert Gu and the legendary Tri Dao. There are many architectures that I expect will break through the Pareto frontier of a mature Transformer, but this one is the first that feels like an actual Vaswani et al. 2017 level advance. Unlimited context, here we come.

Hmm, it seems like I confused the MMMU and MMLU in my original post, despite knowing the difference. I'll edit accordingly.

The MMMU performance seems far more compelling compared to the latter, especially given Dean's methodology of zero-shotting both models.

As someone who is functionally literate, I certainly care more about text prowess, as I presume would most of the people here. But in terms of mundane value for the rest of the world, that will be handy.

Interesting that one of people who had worked on aligning early Gemini said they had trouble aligning it – it burned through RLHF reward models, finding exploits and collapsing into gibberish (imagine using actual RLHF in 2023!). Maybe this has delayed the release, as well as the garden variety safetyism it has made more complex.

Interesting/mildly concerning. I haven't heard any claims of such difficulty in early GPT-4 or Claude, but OAI is probably the best at "alignment" in general, while Anthropic gimps their models to hell.

To be honest I was more excited about the other day's release of Mamba by Albert Gu and the legendary Tri Dao. There are many architectures that I expect will break through the Pareto frontier of a mature Transformer, but this one is the first that feels like an actual Vaswani et al. 2017 level advance. Unlimited context, here we come.

I am the wrong person to comment on such architectural concerns, but if people I respect, such as you and some others, do stress its importance, I'm all for it.

Certainly it seems to me that context windows (along with hallucinations) are the biggest impediments in making LLMs useful for more tasks.

I wonder what the deeper implications for human cognition are. I don't think there are people who can keep 25k words in their working memory, that seems to be much smaller, but we certainly don't usually forget the start of a novella by the time we reach the end. Is there a lot of caching and summarization going on?

At any rate, I hope it beats the annoying reality that 128k and 200k context window models begin to severely underperform, especially for data presented in the middle.

How does it stack up to RWKV?

I wonder what the deeper implications for human cognition are. I don't think there are people who can keep 25k words in their working memory, that seems to be much smaller, but we certainly don't usually forget the start of a novella by the time we reach the end. Is there a lot of caching and summarization going on?

Yes, there is in effect a lot of "caching and summarization" going on -- although that's probably our 2023 ooga-booga, not-quite-wrong way of talking about something else. LLMs really only have their context window and it's feedback as a short-term memory. Which is fine for text translation, but is asinine if you want anything like a thinking engine. Goldfish with a notebook.

We and LLMs can both compress long stories into gists, but the LLMs just forget about it and repeat the work on every iteration. We remember the gists and use them as context on every iteration.

Interesting/mildly concerning.

I think it's a nothingburger because a) the future is cDPO/IPO and not orthodox RLHF anyway (or even more obscure things) and failure modes there will probably be different and b) such «misalignment» results in a behaviorally incoherent model rather than an evil schemer. Reward models are getting hacked by being dragged off-policy, with some weird inputs that are not conductive to strategic world understanding, it's an exploitation of the semiotic nature of language models. But I believe some hay will be made out of it.

Human «context size» is not at all limited to working memory (although our working memory is also large, it's not 5-9 tokens/bits but more like 5-9 «pointers» that can be corresponded to arbitrarily complex cognitive circuits). What we use for context is probably most analogous to constructing on the fly and loading a LoRA in LLMs (or some in-context vector) plus adding embeddings and snippets to some RAG pipeline. It's a mess, but it's orthogonal to the shift from Transformers to SSMs that I expect now. Shane Legg talks of this too:

They don't do things like episodic memory. Humans have what we call episodic memory. We have a working memory, which are things that have happened quite recently, and then we have a cortical memory, things that are sort of being in our cortex, but there's also a system in between, which is episodic memory, which is the hippocampus. It is about learning specific things very, very rapidly. So if you remember some of the things I say to you tomorrow, that'll be your episodic memory hippocampus.
Our models don't really have that kind of thing and we don't really test for that kind of thing. We just sort of try to make the context windows, which is more like working memory, longer and longer to sort of compensate for this.

As for RWKV, I think the latest version is ≤RetNet (though it has good slopes, probably the best in their graph…). Gu&Dao are very explicit in pointing out that a) Mamba the first to even match a Llama-like Transformer without any gimmicks, at the tested scale at least, and b) it does not appreciably benefit from adding Attention layers.

Mamba is the first attention-free model to match the performance of a very strong Transformer recipe (Transformer++) that has now become standard, particularly as the sequence length grows. We note that full results on context length 8k are missing for the RWKV and RetNet baselines, prior strong recurrent models that can also be interpreted as SSMs, due to a lack of efficient implementation leading to out-of-memory or unrealistic computation requirements.

The Mamba-MHA architecture is only slightly better, which is somewhat surprising in light of the fact that many recent works have found that combining (LTI) SSMs with Attention can lead to substantial improvements (Dao, Fu, Saab, et al. 2023; Fathi et al. 2023; Fathullah et al. 2023; Saon, Gupta, and Cui 2023; Zuo et al. 2022).

In the first version of the paper, submitted for peer review, they went even harder:

LongNet (Ding et al., 2023), which claimed to scale to 1B length but only evaluated on length < 100K for actual tasks. Hyena and HyenaDNA (Polietal.,2023;Nguyenetal.,2023),which claimed to leverage up to 1M context, but did not control for computation time. In fact, its claims about efficiency and performance would be largely matched by any of the LTI S4 variants above.

That said, this is all assuming the paper is trustworthy and they compare models trained on identical data. Tri obviously can procure as much compute as needed but I am not sure this happened.

but there's also a system in between, which is episodic memory, which is the hippocampus. It is about learning specific things very, very rapidly. So if you remember some of the things I say to you tomorrow, that'll be your episodic memory hippocampus.

It seems to me that LLMs can't have episodic memory, at least not till they're performing online learning, which nobody is carrying out as far as I'm aware.