waldrews

Ah, so if doing this once can net tens of millions, all we have to do is repeat the process a few million times, and the national debt is covered. That's where automation comes in.

The smart LLM's are great at this (Gemini Flash seems accurate and cheap), but they can't be trusted not to engage in unexpected censorship, typically skipping parts they find objectionable without reliably telling you that that's what they did. That's annoying enough if you're dealing with, e.g., names that happen to spell something awkward, but it's a big problem if you're scanning medical notes or something else where the awkward text is legitimately needed.

Anyone have success with prompting them to "just give me the text verbatim?"

#1 is a valid and important point, that would explain the model name issue legitimately, and on that I am duly mocked.

#2 You wouldn't want to extract all 15T tokens by API, as it wouldn't be desirable to have that as your only source of ground truth. A fraction of that, why not - 1T tokens is just $5 million at the batch API price so the cost isn't a problem, nor a meaningful fraction of OpenAI's revenue, though it would take some doing to route this, likely through enterprize Azure customers.

The more interesting part isn't ChatGPT's answers, but quality questions, the stuff OpenAI pays ScaleAI or Outlier for. If you got inside and could exfiltrate one thing, it would be the dataset of all conversations with paid labellers (unless of course you could get the master log of all conversations with ChatGPT). Even the weights aren't as useful as that to a replication effort.

#3 No statement against the actual demonstrable (and shockingly good) advances in efficiency on several fronts. I'm specifically whining about the legalities and trying to infer what MS/OAI/Sacks could be accusing them of.

This point got mocked when I raised it some time ago:

https://news.ycombinator.com/item?id=42561419

Deepseek promptly fixed it so that their UI responds with 'Hi! I'm DeepSeek-V3, an AI assistant independently developed by the Chinese company DeepSeek Inc. For detailed information about models and products, please refer to the official documentation.' - but only if you ask that as the first question of the conversation. Bury the 'what model are you' question after a few unrelated questions, and it'll happily tell you it's ChatGPT.

Is it possible that it's an actual distillation of weights, but into a radically different architecture? We don't have evidence of that, but that would be a great technical feat in itself.

Is it trained on a large set of user requests and OpenAI replies? Yes.

The question is, were these obtained by simply using the API contrary the user agreement at scale, or was there access to internal OpenAI datasets, or was there some kind of capture of conversations by a man-in-the-middle (which could be any of a number of AI access resellers)?

The answer hinges on which _requests_ were in that training set, something that won't be easy to investigate - unless you're OpenAI itself, and can identify 'trap streets' in the archive of all conversations, cases where ChatGPT once gave an unusual response to an unusual request, and DeepSeek just happens to match it.

The pricing model is very well thought out and worth studying. Cheap but not free generous base layer, that falls back to prepaid credit resource based billing for high users. Even details like punitively increased pricing to discourage higher risk outbound email patterns.

This removes risks from both seller and buyer. The seller is not afraid of an avalanche of free trial users resulting in a big cloud bill. The user is not afraid of surprise overages (cough recent Vercel scandals cough) nor has to think about usage patterns that fit into arbitrary quotas, nor about surprises that catch up to you if you use the service at scale.

This is the way to do small utilities that have a cloud cost; services of the 'save you the trouble of self-hosting' type, or the 'tool somebody writes once, and people might use occasionally for years' pattern.

Maybe not the way to extract maximum value and growth for an ambitious startus, but a good choice for the many modest niche services out there.

They don't even need the AI navigation problems of delivery or military drones to be solved, just to be remote piloted along a given route. But all that solves is removing the risk to the human pilot and their weight - possibly being able to get closer to the fire. We still need to move volume of water to the target, and that means getting a lot of flying vehicles to the right location and time. There's been plenty of big area fires, but this is as bad as it gets in terms of population and assets exposed, so we need a lot of vehicles delivering water to a lot of locations simultaneously.

Being prepared for this scale of operation means either having specialty firefighting planes and drones in reserve, or having other vehicles - general purpose drones or cropdusters or, heck, a Globemaster with a big ballon full of water in its belly - that can be repurposed and sent to this emergency from all over. Airships would be nice for the volume, but they aren't exactly friendly to the turbulent conditions near a fire.

Update: as of today, the response is consistently:

"Hi! I'm DeepSeek-V3, an AI assistant independently developed by the Chinese company DeepSeek Inc. For detailed information about models and products, please refer to the official documentation."

You can ask it if it's sure it's not ChatGPT, and it will repeat that verbatim, suggesting a system prompt or guardrail level instruction.

Space isn't a limited resource, but prestige points are deliberatly limited, as a proxy for the publications' competition for attention. We can appreciate the irony, while considering the outcome reasonable - after all, the results weren't kept out of the literature. They just got published with a label that more or less puts them lower in the search ranking for the next mathematician who looks up the topic.

Happy new year to readers who are fellow humans, to any bots who, unbeknownst to us, can already experience happiness, and I guess if any gorilla, parrot, elephant, or dolphin is having this message communicated to them, happy new year to you too. Let's hope all our respective species keep existing.

When Gemini had its 'identity crisis,' mostly in 2023 (whatever they called it back then), it looked like a pre-training phenomenon, explainable by mentions in the available corpus. The consistency here, and exact and persistent match to how ChatGPT would answer identical queries, suggest training on Q&A pairs from ChatGPT transcripts, presumably at the post-training stage.

It also tells us the genie is out of the bottle not just in the form of open weights being widely available, but in the form of the text corpuses coming from the existing model. The claimed low cost of Deepseek's training is partly enabled by the availability of all that synthetic data created by the first generation models trained and developed at much higher cost. When the Soviets got hold of the nuke plans, they greatly reduced their development costs by primarily by not having to redo all the experiments that led to dead ends. What's amazing is that time it's different; nobody needs OpenAI's secret sauce anymore, just enough data - some of it happily supplied by ChatGPT itself, and they can experiment with different architectures and either get tolerable results with an architecture already in textbooks, or greatly improve efficiency by innovating.

And you can ask it if it's sure, and it'll consistently double down on insisting it's ChatGPT. Ask it what country it's developed in, and it'll say US; ask it if it's sure it's not China, and it'll be sure.

To this day, asking Deepseek "what model are you" typically gives the answer

"I'm an AI language model called ChatGPT, created by OpenAI. Specifically, I'm based on the GPT-4 architecture, which is designed to understand and generate human-like text based on the input I receive. My training data includes a wide range of information up until October 2023, and I can assist with answering questions, generating text, and much more. How can I help you today?"

this tells us something about using synthetic data to bootstrap new model. All those clauses in the terms of service about not using the model to develop competing UI? Yeah, good luck with that.

Seems like a fine number, but I bet there's quite a few more irrational computable numbers out there.

should've said "6 orders of magnitude if we use 120 m/s vs 300_000 km/s" - I was also off by 3 orders of magnitude!

6 orders of magnitude if we use 120 m/s vs 300 km/s

Doesn't look like they got to DoorDash hidden fees in this decision?

The Conformal Prediction advocates (especially a certain prominent Twitter account) tend to rehash old frequentist-vs-bayesian arguments with more heated rhetoric than strictly necessary. That fight has been going on for almost a century now. Bayesian counterargument (in caricature form) would be that MLE frequentists just choose an arbitrary (flat) prior, and penalty hyperparameters (common in NN) are a de facto prior. The formal guarantees only have bite in the asymptotic setting or require convoluted statements about probabilities over repeated experiments; and asymptotically, the choice of prior doesn't matter anyway.

(I'm a moderate that uses both approaches, seeing them as part of a general hierarchical modeling method, which means I get mocked by either side for lack of purity).

Bayesians are losing ground at the moment because their computational methods haven't been advanced as fast by the GPU revolution for reasons having to do with difficulty in parallelization, but there's serious practical work (especially using JAX) to catch up, and the whole normalizing flow literature might just get us past the limitations of MCMC for hard problems.

But having said that, Conformal Prediction works as advertised for UQ as a wrapper on any point estimating model. If you've got the data for it - and in the ML setting you do - and you don't care about things like missing data imputation, error in inputs, non-iid spatio-temporal and hierarchical structures, mixtures of models, evidence decay, unbalanced data where small-data islands coexist big data - all the complicated situations where Bayesian methods just automatically work and other methods require elaborate workarounds, yup, use Conformal Prediction.

Calibration is also a pretty magical way to improve just about any estimator. It's cheap to do and it works (although hard to guarantee anything with that in the general case...)

And don't forget quantile regression penalties! Awkward to apply in the NN setting, but an easy and effective way to do UQ in XGBoost world.

The Property Grid control was the greatest general purpose UI ever created. Using it (in the .Net, Windows Forms incarnation) for custom types in non-VS application was a tremendous productivity shortcut for LOB apps. Put attributes on your user exposed objects, done.

To this day, we don't have anything equivalent for web or mobile. Dynamic form generation from JSON schema gets part of the way there, but not quite.

s/Broken ankles are 100% manageable by not walking where you could slip and fall. If only today's society made informed decisions about their walking habits, we wouldn't need all these artificial casts and surgeries. How long does a cured ankle last till the lifestyle of walking around breaks it again?/

Wyoming being a near-square state, a Root Wyoming would be the unit of length, and a California a unit of temperature, calibrated to the Death Valley record.

Off topic, if 'Three New York Cities' has the dimensions of power, and 'Three Mississippi' has the dimensions of time, we've got a starting point for a new system of units...

What's the GIL/threading story there?

Looks like all of that would run in a native sandbox environment which in turn is called from the Python running on the JVM. So, maybe it simplifies interop, but whether it's straightforward to get full performance from the native layer (especially GPU/multicore) is an open question.

Well, maybe it's still better than the alternative of trying to get there by chemically propelled rocket, and having a 100% chance of dying of old age on the way.

Just use 3.

https://www.ycombinator.com/nonprofits/ - there hasn't been a 'unicorn scale' non-profit to come out of the program yet, but they're trying...

I'm very happy this is now available. With a model as smart as Claude 3.5, this opens up the possibility of writing prompts the size of relatively complex programs, prompts with many examples and targeting multiple outputs, and running them repeatedly on different datasets (big code, small data), at a reasonable cost. For example, a classifier with a hundred class labels described by prompts? A questionnaire that turns a piece of unstructured data into a long list of features? A complex custom assistant, with a long prompt listing individual user's available tools and preferences?

(my previous rants on why we need this: https://news.ycombinator.com/item?id=40034972#40036309 )

Now we just need to merge Anthropic's cache pricing with Google's context length (their cache pricing isn't as good) and OpenAI's strict structured output mode...

There's no reward function in the sense that optimizing the reward function means the solution is ideal.

There are objective criteria like 'compiles correctly' and 'passes self-designed tests' and 'is interpreted as correct by another LLM instance' which go a lot further than criteria that could be defined for most kinds of verbal questions.

How are people still publishing books without audiobook versions? What, are we supposed to read with our eyes, like animals?