I skimmed the doc and my impression is that your second listed interpretation -- DeepSeek is pausing investment because of a leak -- is the more correct one.
There's quite a bit of confidential information in the doc about the company and how it's positioning itself going forward to compete with US labs. I'd imagine they're not happy at all with this being leaked and are withholding investment as a punitive measure.
Not to mention the other interpretation seems illogical -- why would you pause fundraising if your perception was that you lacked resources compared to your competitors?
I've found breathing exercises to be effective for the duration of the exercise, but I'm more interested in the possibility of training myself to adjust my respiration patterns over sustained durations. Would it be beneficial -- or even possible at all -- to adjust my body's default/subconscious breathing patterns to match those mentioned in the article?
Tangentially related, are there any wearable devices that allow for high resolution respiration monitoring? I'm imagining some measurement of lung expansion over time (probably at least 10 Hz) so that I can quantify the deepness/shallowness of my breaths as well as the phase of inhalation/exhalation cycles.
Kind of a backwards take, both in your interpretation of this announcement and the company in general.
The point of this announcement is to draw attention to the fact that the currently hyped topic is what they have been working on since their inception. If anything, it gives off a Schmidhuber-esque 'actually it was me who invented that' vibe. But trying to retroactively claim credit for the hype is nowhere near the same as following the hype.
As for your impression that the company is more generally hype-chasing, I'm really not sure how you would come to that conclusion. At the time of their founding, chatbots were the hype on the product side and model scaling was the hype on the research side -- topics they have largely eschewed. They instead were founded with a focus on evolutionary and collective intelligence and have maintained a fairly cohesive research direction ever since.
I can assure you that a fully deterministic and equally effective claude is possible to build. And yes, that would mean identical prompts would yield 100% identical output 100% of the time. It would still make the occasional logical or factual error, but it would do so deterministically. Would this solve any of the problems with building reliable programs using LLMs?
I see what you're getting at, but determinism isn't the right word either. LLMs are fundamentally deterministic -- they are pure functions which output text as a function of the input text and the network parameters[1]. Depending on your views on free will, it could be effectively argued that humans are deterministic as well.
The concept you're touching on is the idea that LLMs (and humans) are functions which are inscrutable. Their behavior cannot be distilled into a series of logical steps that you can fit in your head, there are no invariants which neatly decompose their complexity into a few interpretable states, and the input and output spaces are unstructured, ambiguous, underspecified, and essentially infinite. This makes them just about impossible to reason about or compose using the same strategies and analysis we apply to traditional programs.
[1] Optionally, they can take in a source of entropy to add nondeterminism, but this is not essential. If LLM providers all fixed their prng seeds to a static value, hardly anyone would notice. I can't imagine there are many workflows which feed an LLM the exact same prompt multiple times and rely on the output having some statistical distribution. In fact, even if you wanted this you may just end up getting a cached response.
The guy writing a thumbnail pipeline isn't getting petabytes (exabytes?) of storage to cache all videos from the past week in their entirety. If this quantity of data is being stored, it's being stored deliberately and at significant cost.
The other explanations here don't explain the long delay between the start of the investigation and the release of the footage.
Yes, storing customer data is what we'd expect from Google and yes, the FBI can coerce Google to provide this data for their investigations.
But it does not take a week for Google to find a file on their servers.
My hunch is that Google initially tried to play dumb to avoid compliance, as to not reveal they do in fact retain customer data.
They had a plausible excuse as well -- the owner had no subscription so they don't store the data -- and took a gamble that this explanation would suffice until the situation resolved itself.
I suspect that authorities initially took Google's excuse at face value, since they parroted this explanation to the public as well.
As pressure mounted on authorities to make some headway on the case, they likely formally exercised whatever legal mechanisms they have at their disposal to force Google's hand, and only then was the footage released.
The implication that OpenAI is a YC company in the same sense as the other listed companies is somewhere between misleading and dishonest. Even more distasteful to show founding teams for all the others, then just Sam for OpenAI.
Of all Schmidhuber's credit-attribution grievances, this is the one I am most sympathetic to. I think if he spent less time remarking on how other people didn't actually invent things (e.g. Hinton and backprop, LeCun and CNNs, etc.) or making tenuous arguments about how modern techniques are really just instances of some idea he briefly explored decades ago (GANs, attention), and instead just focused on how this single line of research (namely, gradient flow and training dynamics in deep neural networks) laid the foundation for modern deep learning, he'd have a much better reputation and probably a Turing award. That said, I do respect the extent to which he continues his credit-attribution crusade even to his own reputational detriment.
My personal experience is that the cost of enduring a negative stimulus is not simply a function of the magnitude of the negative stimulus, but rather the magnitude of the negative stimulus in relation to the magnitude of all other concurrent negative stimuli. This study controls the environment so that a single negative stimulus is isolated and additional external negative stimuli are minimized, but it cannot control for the fact that a depressed person also endures a constant barrage of negative stimuli which are generated internally (hopelessness, exhaustion, fear, self-doubt, etc). The magnitude of these internally generated negative stimuli is likely much larger than that of the aversive external stimulus used in this study, so it seems reasonable that the marginal relief obtained by avoiding the external stimulus may be perceived as relatively negligible, or at least diminished to the point that the cost of avoiding is greater than the cost of enduring.
Key-based attention is not attributable to the Transformer paper. First paper I can find where keys, queries, and values are distinct matrices is https://arxiv.org/abs/1703.03906, described at the end of section 2. The authors of the Transformer paper are very clear in how they describe their contribution to the attention formulation, writing "Dot-product attention is identical to our algorithm, except for the scaling factor". I think it's fair to state that multi-head is the paper's only substantial contribution to the design of attention mechanisms.
I think you're overestimating the degree to which this type of research is motivated by big-picture, top-down thinking. In reality, it's a bunch of empirically-driven, in-the-weeds experiments that guide a very local search in a intractably large search space. I can just about guarantee the process went something like this:
- The authors begin with an architecture similar to the current SOTA, which was a mix of recurrent layers and attention
- The authors realize that they can replace some of the recurrent layers with attention layers, and performance is equal or better. It's also way faster, so they try to replace as many recurrent layers as possible.
- They realize that if they remove all the recurrent layers, the model sucks. They're smart people and they quickly realize this is because the attention-only model is invariant to sequence order. They add positional encodings to compensate for this.
- They keep iterating on the architecture design, incorporating best-practices from the computer vision community such as normalization and residual connections, resulting in the now-famous Transformer block.
At no point is any stroke of genius required to get from the prior SOTA to the Transformer. It's the type of discovery that follows so naturally from an empirically-driven approach to research that it feels all but inevitable.
The scientific impact of the transformer paper is large, but in my opinion the novelty is vastly overstated. The primary novelty is adapting the (already existing) dot-product attention mechanism to be multi-headed. And frankly, the single-head -> multi-head evolution wasn't particularly novel -- it's the same trick the computer vision community applied to convolutions 5 years earlier, yielding the widely-adopted grouped convolution. The lasting contribution of the Transformer paper is really just ordering the existing architectural primitives (attention layers, feedforward layers, normalization, residuals) in a nice, reusable block. In my opinion, the most impactful contributions in the lineage of modern attention-based LLMs are the introduction of dot-product attention (Bahdanau et al, 2015) and the first attention-based sequence-to-sequence model (Graves, 2013). Both of these are from academic labs.
As a side note, a similar phenomenon occurred with the Adam optimizer, where the ratio of public/scientific attribution to novelty is disproportionately large (the Adam optimizer is very minor modification of the RMSProp + momentum optimization algorithm presented in the same Graves, 2013 paper mentioned above)
Tensorflow is a lot like IBM -- it deserves praise not because it's great in its current state, but for its contributions towards advancing the broader technological front to where it is today. Tensorflow walked so JAX could run, so to speak. Frankly, I don't really draw much of a distinction between the two frameworks since I really just use them as lightweight XLA wrappers.
I've seen and ignored a lot of "pytorch good, tensorflow bad" takes in my time, but this is so egregiously wrong I can't help but chime in. Facilitating graph-level optimizations has been one of the most central tenets of tensorflow's design philosophy since its inception. The XLA compiler was designed in close collaboration with the tensorflow team and was available in the tensorflow API as far back as 2017. It's not an exaggeration to say that pytorch is 5+ years behind on this front. Before anyone invokes the words "pythonic" or "ergonomic", I'd like to note that the tensorflow 2 API for compilation is nearly identical to torch.compile.
I think that while researchers would agree that the empirical success of deep learning has been remarkable, they would still agree that the language used here -- "an algorithm that could really, truly learn any distribution of data (or really, the underlying “rules” that produce any distribution of data)" -- is an overly strong characterization, to the point that it is no longer accurate. A hash function is a good example of a generating process which NN + SGD will not learn with any degree of generalization. If you trained GPT4 on an infinite dataset of strings and their corresponding hashes, it would simply saturate its 100 billion+ parameters with something akin to a compressed lookup table of input/output pairs, despite the true generating process being a program that could be expressed in less than a kilobyte. On unseen data, it would be no better than a uniform prior over hashes. Anyways, my point is that people knowledgable in the field would have far more tempered takes on the practical limits of deep learning, and would reserve the absolute framing used here for claims that have been proven formally.
> humanity discovered an algorithm that could really, truly learn any distribution of data (or really, the underlying “rules” that produce any distribution of data)
He's hand-waving around the idea presented in the Universal Approximation Theorem, but he's mangled it to the point of falsehood by conflating representation and learning. Just because we can parameterize an arbitrarily flexible class of distributions doesn't mean we have an algorithm to learn the optimal set of parameters. He digs an even deeper hole by claiming that this algorithm actually learns 'the underlying “rules” that produce any distribution of data', which is essentially a totally unfounded assertion that the functions learned by neural nets will generalize is some particular manner.
> I find that no matter how much time I spend thinking about this, I can never really internalize how consequential it is.
If you think the Universal Approximation Theorem is this profound, you haven't understood it. It's about as profound as the notion that you can approximate a polynomial by splicing together an infinite number of piecewise linear functions.
Is batched inference for LLMs memory bound? My understanding is that sufficiently large batched matmuls will be compute bound and flash attention has mostly removed the memory bottleneck in the attention computation. If so, the value proposition here -- as well as with other memorymaxxing startups like Groq -- is primarily on the latency side of things. Though my personal impression is that latency isn't really a huge issue right now, especially for text. Even OpenAI's voice models are (purportedly) able to be served with a latency which is a low multiple of network latency, and I expect there is room for improvement here as this is essentially the first generation of real-time voice LLMs.
There's quite a bit of confidential information in the doc about the company and how it's positioning itself going forward to compete with US labs. I'd imagine they're not happy at all with this being leaked and are withholding investment as a punitive measure.
Not to mention the other interpretation seems illogical -- why would you pause fundraising if your perception was that you lacked resources compared to your competitors?