airgapstopgap

Can you point to anyone other than yourself who calls Indonesians black? Because I think otherwise it's not worthwhile discussing categorization and measurement in good faith with you.

Since you're here: have you considered moving to other, better generalist base models in the future? Particularly Deepseek or Mixtrals. Natural language foundation is important for reasoning. Codellama is very much a compromise, it has lost some NLP abilities from continued pretraining on code.

Note that we have no reason to believe that the underlying LLM inference process has suffered any setbacks. Obviously it has generated some logits. But the question is how is OpenAI server configured and what inference optimization tricks they're using.

This is not so surprising if you consider the fact that finetuning is extremely sparse and barely imparts any new knowledge to the model. The paper "Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch"[1] made this clear:

> We initially demonstrate that SFT LM (either encoder- or decoder-based) always tends to acquire excessively redundant delta parameters. To be specific, we present DARE, which randomly resets some delta parameters to zeros based on a drop rate p and subsequently scales the remaining parameters by a factor of 1/(1 − p). Despite its simplicity, with the assistance of DARE, when the LM model parameters reach 70 billion, we can eliminate up to 99% delta parameters with minimal impact on model performance (see Figure 1(a)). The more parameters the LM has, the larger p it can tolerate. This discovery suggests that SFT LM indeed learns a multitude of low-rank structures akin to LoRA [25]

Insofar as those adaptations are mostly distinct, you can just preserve both sets and that's what explains successes of merging, I guess.

1. https://arxiv.org/abs/2311.03099

Intel aims to.

The original paper by Shazeer suffices. What you are saying is in theory possible to do and may have been done in practice here, but in the general case MoE is trained from scratch and specializations of layers which develop are not products of some design choice.

Mistral-small explicitly has inference costs of a 12.9b, but more than that, it's probably ran with batch size of 32 or higher. They'll worry more about offsetting training costs than about this.

Here's how it works in reality:

https://docs.mystic.ai/docs/mistral-ai-7b-vllm-fast-inferenc...

> It's not even close to a 45B model. They trained 8 different fine-tunes on the same base model. This means the 8 models differ only by a couple of layers and share the rest of their layers.

No, Mixture-of-Experts is not stacking finetunes of the same base model.

> today we have no architecture or training methodology which would allow it to be possible.

We clearly see that Mistral-7B is in some important, representative respects (eg coding) superior to Falcon-180B, and superior across the board to stuff like OPT-175B or Bloom-175B.

"Well trained" is relative. Models are, overwhelmingly, functions of their data, not just scale and architecture. Better data allows for yet-unknown performance jumps, and data curation techniques are a closely-guarded secret. I have no doubt that a 7B beating our best 60-70Bs is possible already, eg using something like Phi methods for data and more powerful architectures like some variation of universal transformer.

I do not even think any of this has much of impact on AGI timelines. Human brain cells are not a superior substrate for computing "intelligence". They just are what they are; individual cells can somewhat meaningfully "want" stuff and be quasi-agents unto themselves, they do much more than integrate and fire. Weights in an ANN are purely terms in an equation without any inner process or content.

Comments like this are incredibly grating. You condescend to the interlocutor for making a mistake which only exists in your own mistaken world model. Your confidence that neurons and ANN weights and «pulleys and gears» are all equivalent because there is, in theory, an intention to instantiate some computation, and to think otherwise is tantamount to belief in magic and broken causality, is just confused and born out of perusing popular-scientific materials instead of relying on scientific literature or hands-on experience.

> The fire because of input.

No they do not fire because of input, they modulate their firing probability based on input, and there are different modalities of input with different effects. Neurons are self-contained biological units (descended, let me remind you, from standalone unicellular organisms, just like the rest of our cells), which actually have an independently developing internal state and even metabolic needs; they are not merely a system of logic gates even if you can approximate their role with a system of equations or an ANN. This is very different, mechanistically and teleologically. Hell, even spiking ANNs would be substantially different from currently dominant models.

> So what, magic ? a soul ? If the brain is computing then the substrate is entirely irrelevant

Stop dumbing down complex arguments to some low-status culture war opinion you find it easy to dunk on.

…ETH Zurich is an illustrious research university that often cooperates with Deepmind and other hyped groups, they're right there at the frontier too, and have been for a very long time. They don't have massive training runs on their own but pound for pound I'd say they have better papers.

> murderous tendencies lurking beneath the surface

…Where is that "beneath the surface"? Do you imagine a transformer has "thoughts" not dedicated to producing outputs? What is with all these illiterate anthropomorphic speculations where an LLM is construed as a human who is being taught to talk in some manner but otherwise has full internal freedom?

> there is a possibility that for things like AI, with extra time comes the ability to better understand and build those defenses before they're needed.

Or not, and damaging wrongheaded ideas will become a self-reinforcing (because safety! humanity is at stake!) orthodoxy, leaving us completely butt-naked before actual risks once somebody makes a sudden clandestine breakthrough.

https://bounded-regret.ghost.io/ai-pause-will-likely-backfir...

> We don’t need to speculate about what would happen to AI alignment research during a pause—we can look at the historical record. Before the launch of GPT-3 in 2020, the alignment community had nothing even remotely like a general intelligence to empirically study, and spent its time doing theoretical research, engaging in philosophical arguments on LessWrong, and occasionally performing toy experiments in reinforcement learning.

> The Machine Intelligence Research Institute (MIRI), which was at the forefront of theoretical AI safety research during this period, has since admitted that its efforts have utterly failed. Other agendas, such as “assistance games”, are still being actively pursued but have not been significantly integrated into modern deep learning systems— see Rohin Shah’s review here, as well as Alex Turner’s comments here. Finally, Nick Bostrom’s argument in Superintelligence, that value specification is the fundamental challenge to safety, seems dubious in light of LLM's ability to perform commonsense reasoning.[2]

> At best, these theory-first efforts did very little to improve our understanding of how to align powerful AI. And they may have been net negative, insofar as they propagated a variety of actively misleading ways of thinking both among alignment researchers and the broader public. Some examples include the now-debunked analogy from evolution, the false distinction between “inner” and “outer” alignment, and the idea that AIs will be rigid utility maximizing consequentialists (here, here, and here).

> During an AI pause, I expect alignment research would enter another “winter” in which progress stalls, and plausible-sounding-but-false speculations become entrenched as orthodoxy without empirical evidence to falsify them. While some good work would of course get done, it’s not clear that the field would be better off as a whole. And even if a pause would be net positive for alignment research, it would likely be net negative for humanity’s future all things considered, due to the pause’s various unintended consequences. We’ll look at that in detail in the final section of the essay.

Long-context tasks are not really the true gap between LLaMA and GPT series, but important result.

Being authors of LLaMA is sufficient to argue they know how to train LLaMAs.

Interested about your logic, what did you like about pre-LLM AGI? The "maximize utility function at any cost" feature? The single-minded focus on beating people in games?

It's quite terrifying how, as we've chosen an apparently very easy path to bake our preferences and quirks into intelligent systems, people became very "responsible" and concerned for survival of human race, parroting alarmist rhetoric that precedes not only LLMs, but even RL successes of early Deepmind and just cites vague shower thoughts of Bostrom and such non-technical ilk. Say what you want about LLMs but there's zero credible reason to perceive them as a more risky approach!

Provable safety (not to confuse with security as in normal discussion of vulnerabilities) for general intelligence is a pipe dream because, putting things simply, undesirable reasoning in full generality is not a meaningful class of computations. The end result of this line of thinking is centralization of AI development in a state-approved and military-associated facility.

> Why would governments promote provably secure systems?

Promote? The state demonstrably wants provably secure systems for themselves, in the military but also in the civilian sphere, see Matrix/Element, see DoD, see massive state interest in cryptography. This is an incredibly disingenuous argument, you talk as if people discuss tuning a generalized Safety Dial without any distinctions down the line.

Tegmark's thinking here is extremely shallow, discards the costs (opportunity costs and risks of stable dystopia) associated with this grandiose global project of dubious feasibility, and indeed I suspect he does not so much believe his own arguments as he generally prefers global technocratic regulation "for good measure". We don't really have good reasons to expect uncontrollable AGI in a way they posit (increasingly less as we advance down this path of ML model scaling), but we are well acquainted with unaccountable human power structures.

I wonder if you have enough self-awareness to notice why your behavior here might be considered bizarre. No, people who point out that your government routinely and brazenly backdoors equipment and software everyone uses (or rather, has forced upon them to use), while Chinese actions of this sort are evidenced to a lesser extent, and so the onus in justifying the use of that country's products is on the US at least as much as on China, are mostly not "wumao", nor whataboutists, nor anything of the sort. They're making an entirely sound argument an unbiased person would make given the context, you're just using those mind-killing political notions to dismiss a topic that offends your patriotic sensibilities.

Wumaos are low-paid grunts and sincere idiots who disingenuously downvote, report and post irrelevant nonsense regarding racist imperialist AmeriKKKa or legitimate Chinese clay/territorial waters/6000 years of peaceful history. This is very easy to see. You're free to suspect any interlocutor as being one, of course, but if that's your only retort, you'd do better not stooping to the level of an undeniable propagandist and instead conceding the object-level issue – or keeping silent.

Llama-1-33B was trained on 40% more tokens than LLama-1-13B; this explained some of the disparity. This time around they both have the same data scale (2T pretraining + 500B code finetune), but 34B is also using GQA which is slightly more noisy than MHA. Furthermore, there have been some weird indications in the original LLama-2 paper that 34B base model is something… even more special, it's been trained on a separate internal cluster with undervolted/underclocked GPUs (though this in itself can't hurt training results), its scores are below expectations, it's been less "aligned". Here, Code-Llama-Instruct-13B is superior to 34B on HumanEval@1. So yes, it's desirable but I wouldn't get my hopes up.

> Linux and Mac

> Coming soon ...

Ah well. Hopefully it is soon. Also, on behalf of all Apple Silicon Mac users, would be nice if the author looked into implementing Metal FlashAttention [1].

1. https://github.com/philipturner/metal-flash-attention

This is an incredible achievement but there are strong reasons to suspect that stellarators are not and will never be plausible candidates for energy generation. For some more experimental or perhaps military tasks, it's viable.

Do you not consider that Huawei "executive's" detention (actual makes for a similar case against Canada? It was a purely political move, Meng Wanzhou was detained on grounds of a broader anti-Huawei campaign by the US.

You are frustrated and this makes you act in a deliberately obtuse manner.

There is a world of difference between "anyone who has worked with the guy" and "has worked with the guy + has hundreds of comments on HN identifying career track over the last few years". The former grants each suspect plausible deniability, while the latter pinpoints the true author.

> Anonymity is not a realistic goal.

It obviously is for a throwaway account. Time to reread the classic

https://terrytao.wordpress.com/about/anonymity-and-the-inter...

It is entirely believable that a person with substantial trace on HN and ties in the field would rather create a throwaway than post such remark under his main account.

No, there's no YB and they propose an entirely novel mechanism for how generic metals (a whole host of possible combinations) can achieve superconductivity in these conditions. At least check out the formula or open the link.

> modified lead-apatite (LK-99) structure

> The superconductivity of LK-99 originates from minute structural distortion by a slight volume shrinkage (0.48 %), not by external factors such as temperature and pressure. The shrinkage is caused by Cu2+ substitution of Pb2+ (2) ions in the insulating network of Pb(2)-phosphate and it generates the stress.

> Pb10(PO4)6O

It's just Lead, Phosphorus and Oxygen all the way.

(That said I don't believe it works)

> The same goes for using T5-XXL

Is this still true in 2023? Sure, back in the dark ages it seemed like a 860M model is just about the limit for a regular consumer, but I don't see why we wouldn't be able to use quantized encoders; and even 30B LLMs run okay on Macbooks now.

Diffusion is more parameter-efficient and you quickly saturate the target fidelity, especially with some refiner cascade. It's a solved problem. You do not need more than maybe 4B total. Images are far more redundant than text.

In fact, most interesting papers since Imagen show that you get more mileage out of scaling the text encoder part, which is, of course, a Transformer. This is what drives accuracy, text rendering, compositionality, parsing edge cases. In SD 1.5 the text encoder part (CLIP ViT-L/14) takes a measly 123M parameters.[1] In Imagen, it was T5-XXL with 4.6B [2]. I am interested in someone trying to use a really strong encoder baseline – maybe from a UL2-20B – to push this tactic further.

Seeing as you can throw out diffusion altogether and synthesize images with transformers [3], there is no reason to prioritize the diffusion part as such.

1. https://forums.fast.ai/t/stable-diffusion-parameter-budget-a...

2. https://arxiv.org/abs/2205.11487

3. https://arxiv.org/abs/2301.00704

No, it makes sense to secure engagement with the most expensive implementation and then cut costs, this kind of stuff is pervasive in the industry. Besides, we have Brockman on record saying that they do "a lot of quantization"[1][2] so it's not paranoia to suspect other optimization schemes when there's a clear performance drop, which they have also denied a few times.

1. https://chat.openai.com/share/44a0c5b6-c629-470a-992f-8cdbbe...

2. https://www.youtube.com/watch?v=_hpuPi7YZX8