HN user
LoRA? The parameter-efficient fine-tuning method published 2 years before Llama and already actively used by researchers?
RoPE? The position encoding method published 2 years before Llama and already in models such as GPT-J-6B?
DPO, a method whose paper had no experiments with Llama?
QLoRA? The third in a series of quantization works by Tim Dettmers, the first two of which pre-dated Llama?
If your starting position is already that Sam Altman lies about everything that doesn't fit your preconceived positions, that doesn't seem like a very useful meaningful position to update.
Surely if OpenAI had insisted upon the same things that Anthropic had, the government would not have signed this agreement.
But they did.
"Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement."
that they need to rig their elections against themselves to get dissenting voices
I don't believe this is true. If you're talking about Non-Constituency Members of Parliament, they are consolation prizes given to best losers, and there are many things they cannot vote on. Moreover, the ruling party almost never lifts the party whip, i.e. members of the party CANNOT vote against the party line (without being kicked out of the party, which results in them being kicked out of parliament). In other words, since the ruling party already has a majority, any opposing votes literally do not matter.
If you aren't talking about the NCMP scheme, then I do not know what you're talking about, as the ruling party does institute policies that are beneficial for the incumbent party.
My statement was
a (fine-tuned) base Transformer model just trivially blowing everything else out of the water
"Attention is All You Need" was a Transformer model trained specifically for translation, blowing all other translation models out of the water. It was not fine-tuned for tasks other than what the model was trained from scratch for.
GPT-1/BERT were significant because they showed that you can pretrain one base model and use it for "everything".
GPT-1 wasn't used as a zero-shot text generator; that wasn't why it was impressive. The way GPT-1 was used was as a base model to be fine-tuned on downstream tasks. It was the first case of a (fine-tuned) base Transformer model just trivially blowing everything else out of the water. Before this, people were coming up with bespoke systems for different tasks (a simple example is that for SQuAD a passage-question-answering task, people would have an LSTM to read the passage and another LSTM to read the question, because of course those are different sub-tasks with different requirements and should have different sub-models). One GPT-1 came out, you just dumped all the text into the context, YOLO fine-tuned it, and trivially got state on the art on the task. On EVERY NLP task.
Overnight, GPT-1 single-handedly upset the whole field. It was somewhat overshadowed by BERT and T5 models that came out very shortly after, which tended to perform even better on the pretrain-and-finetune format. Nevertheless, the success of GPT-1 definitely already warrants scaling up the approach.
A better question is how OpenAI decided to scale GPT-2 to GPT-3. It was an awkward in-between model. It generated better text for sure, but the zero-shot performance reported in the paper, while neat, was not great at all. On the flip side, its fine-tuned task performance paled compared to much smaller encoder-only Transformers. (The answer is: scaling laws allowed for predictable increases in performance.)
Because the author is artifically shrinking the scope of one thing (prompt engineering) to make its replacement look better (context engineering).
Never mind that prompt engineering goes back to pure LLMs before ChatGPT was released (i.e. before the conversation paradigm was even the dominant one for LLMs), and includes anything from few-shot prompting (including question-answer pairs), providing tool definitions and examples, retrieval augmented generation, and conversation history manipulation. In academic writing, LLMs are often defined as a distribution P(y|x) where X is not infrequently referred to as the prompt. In other words, anything that comes before the output is considered the prompt.
But if you narrow the definition of "prompt" down to "user instruction", then you get to ignore all the work that's come before and talk up the new thing.
I believe the above post was highlighting that as a misconception young people may have, not saying it is the case.
Two points to consider, one against and one for.
1) It's a small island, but it's also a major trading port. Which means its whole economy is already geared towards importing food from neighboring countries.
2) On the other hand: no domestic industry to disrupt! No domestic farming groups lobbying against meat substitutes, which may push research/distribution furhter along.
The long story short is you are technically correct but in practice things are a little different. There are 2 factors to consider here:
1. Model Capability
You are right that mechanically, input and output tokens in a standard decoder Transformer are "the same". A 32K context should mean you can have 1 input tokens and 32K output tokens (you actually get 1 bonus token), or 32K input tokens and 1 output token,
However, if you feed an LM "too much" of its own input (read: have too long an output length), it starts to go off the rails, empirically. The word "too much" is doing some work here: it's a balance of both (1) LLM labs having data that covers that many output tokens in an example and (2) LLMs labs having empirical tests to have confidence that the model won't reasonably go off the rails within some output limit. (Note, this isn't pretraining but the instruction tuning/RLHF after, so you don't just get examples for free)
In short, labs will often train a model targeting an output context length, and put out an offering based on that.
2. Infrastructure
While mathematically having the model read external input and its own output are the same, the infrastructure is wildly different. This is one of the first things you learn when deploying these models: you basically have a different stack for "encoding" and "decoding" (using those terms loosely. This is after all still a decoder only model). This means you need to set max lengths for both encoding and decoding separately.
So, after a long time of optimizing both the implementation and length hyperparameters (or just winging it), the lab will decide "we have a good implementation for up to 31K input and 1k output" and then go from there. If they wanted to change that, there's a bunch of infrastructure work involved. And because of the economies of batching, you want many inputs to have as close to the same lengths as possible, so you want to offer fewer configurations (some of this bucketing may be performed hidden from the user). Anyway, this is why it may become uneconomical to offer a model at a given length configuration (input or output) after some time.
You could easily make the other argument: As a professor of ethics she studies many different ethical systems, including ones that are not mainstream. This means that she can more easily find some ethical system under which a given action is considered ethical.
The "ethics expert = more ethical" connection has never held up and mainly serves as a gotcha.
Is it stated somewhere that Radford was inspired by that blog post?
It is no coincidence that EleutherAI named their pretraining dataset "the Pile"
The Pythia models have all the training data, code, and configurations available.
EleutherAI as well.
This arguments feels like it's trying to be an inch too smart.
Consider the following: Amazon isn't really an online retail company; it doesn't really sell goods to the consumer. What it does is use "goods" that it delivers as a loss leader to get people to click on buttons to give Amazon money.
Starbucks is a coffee company, that introduces an optional extra step of gift cards for other consumer convenience reasons. The fact that there is a comically large amount of money held in the gift cards system is just that: a slightly comical fact.
the RoPE embeddings in Code Llama were designed for this.
The RoPE embeddings were not "designed" for that. The original RoPE was not designed with length extrapolation in mind. Subsequent tweaks to extrapolate RoPE (e.g. position interpolation) are post-hoc tweaks (with optional tuning) to an entirely vanilla RoPE implementation.
BERT was on arXiv before being peer reviewed. As were T5, BART, LLaMA, OPT and GPT-NeoX-20B. The Pile and FLAN were also on arXiv before being peer reviewed. Of course, the original Transformer paper was also on arXiv before being peer reviewed.
Being on arXiv before being peer reviewed is not the or even a problem.
Makes sense! But expensive...
But what would they be calling out?
If industry groups want to run a training run based on the configurations of a well-performing model, I don't see anything wrong with that. Now, if they were to claim that what they are doing is somehow "optimal", then there would be something to criticize.
Yep I understood that you were using it informally, just trying to keep things informative for other folks reading too.
I want to jump in and correct your usage of "LLaMA Laws" (even you are using it informally, but I just want to clarify).
There is no "LLaMA scaling law". There are a set of LLaMA training configurations.
Scaling laws describe the relationship between training compute, data, and expected loss (performance). Kaplan et al., estimated one set of laws, and the Chinchilla folks refined that estimate (mainly improving it by adjusting the learning rate schedule).
The LLaMA papers do not posit any new law nor contradict any prior one. They chose a specific training configuration that still abide by the scaling laws but with a different goal in mind.
(Put another way: a scaling law doesn't tell you what configuration to train on. It tells you what to expect given a configuration, but you're free to decide on whatever configuration you want.)
If you want a speedrun explanation for how we get to "2": In the limit of model scaling, context size doesn't matter (yes, forget about the quadratic attention), most of the compute is in the linear layers, which boil down to matrix multiplies. Consider a single matrix of size [T,d] multiplied by weight of size [d,d], the compute needed for a matrix multiplication is approximately 2Td^2 (2 coming from multiply + add). Swap T out with D for your whole dataset in tokens, d^2 is the number of parameters in a single linear layer so scale up your model to P, and you've got 2PD.
Even shorter: The 2 comes from the multiply-add
It's actually even less remarkable than that. It was an experiment in having a limited release, to shift the field toward a different release convention.
Nearly a year ago we wrote in the OpenAI Charter: “we expect that safety and security concerns will reduce our traditional publishing in the future, while increasing the importance of sharing safety, policy, and standards research,” and we see this current work as potentially representing the early beginnings of such concerns, which we expect may grow over time.
This decision, as well as our discussion of it, is an experiment: while we are not sure that it is the right decision today, we believe that the AI community will eventually need to tackle the issue of publication norms in a thoughtful way in certain research areas.
We will further publicly discuss this strategy in six months.
While MoE-LoRAs are exciting in themselves, they are a very different pitch from full on MoEs. If the idea behind MoEs is that you want completely separate layers to handle different parts of the input/computation, then it is unlikely that you can get away with low-rank tweaks to an existing linear layer.
Another example that I read about once and have never been able to verify (or it may be completely made up) is that the because the Chinese invented porcelain first (which was more sturdy than glass or something) they never bothered with glass, which meant they missed out on all the cool astronomical discoveries (which then has implications on their development of mathematics and physics).
Again, no idea if there is any validity to this or just something completely made up.
As a researcher in the field, I agree with this characterization.
I think it's more accurate the say that GPT and then BERT massively popularized and simplified the idea/approach. Prior to ULMFiT/GPT/BERT, fine-tuning usually meant freezing most of a model and tuning a small layer on top of it. (ELMo also fits somewhere in here, being a kind of halfway step). ULMFiT was a relatively lesser known work but to my knowledge one of the first to do 1) LM pretraining and 2) fine-tuning all the layers, albeit with some complexity (gradual unfreezing/different learning rates for layers).
GPT simplified this massively by simply tuning all weights: no nuance about it. (Also added a classifier on top). BERT took this a step further, benefiting from the larger size of BERT-large and bidirectional attention, which works really well on the NLU datasets of the time along with a classifier head. (It's not until T5 that seq2seq for general tasks became prominent again.) The idea that you tune all the weights of what was then considered a massive model was considered somewhat excessive at the time, but BERT so massively dominated every other architecture and approach at the time that everyone switched over to it. (Adapters (alongside PAL), one of the first parameter-efficient tuning methods in the Transformer era of NLP, came out shortly after.)
I don't think this "burn it all down" mentality is helpful.
Presumably the mods are protesting because they want a good Reddit/subreddit experience, and the new policies are hurting that. How is handing power over to someone who is supportive of Reddit's new policies helping that?
This really should be it. Open up but slowly deteriorate the experience.
Here's a concrete proposal: Set automod to delete any post/comment longer than 20 characters.
This makes discussion virtually impossible, and starts to impact the quality of Reddit posts. With enough subs cooperating (e.g. the subs that went private over the blackout), it'd already be making a serious impact on the usage experience of Reddit. Furthermore, this policy can be enacted indefinitely without stopping the "regular operation" of Reddit (i.e. mods can sit on this policy for as long as they need to, as opposed to the blackout which even from the start was planned with a time limit.)
It depends on the quantization method, but yes some of the most commonly used ones are extremely slow.