xianshou

Aha! Finally, a perfect spiritual complement to the Gervais principle:

https://www.ribbonfarm.com/2009/10/07/the-gervais-principle-...

According to Rao, every company survives by blending some combination of the following populations, varying by stage of lifecycle:

- Sociopath = people who know the game and play it to win

- Clueless = people who buy and spread the story the company is selling

- Loser = people who accept the wage bargain, generally exchanging low devotion for minimal advancement

This maps nicely to OP's quadrants:

- Sociopath = grinder-grifter

- Clueless = grinder-believer

- Loser = coaster-grifter

- (Coaster-believers get fired.)

From the actual huggingface site - seems like API access leaked and a few mini-videos generated over the ~3 hours the leak was up (https://huggingface.co/spaces/PR-Puppets/PR-Puppet-Sora):

==================================================

┌∩┐(◣◢)┌∩┐ DEAR CORPORATE AI OVERLORDS ┌∩┐(◣◢)┌∩┐ If this letter resonates with you add your signature here.

We received access to Sora with the promise to be early testers, red teamers and creative partners. However, we believe instead we are being lured into "art washing" to tell the world that Sora is a useful tool for artists.

ARTISTS ARE NOT YOUR UNPAID R&D we are not your: free bug testers, PR puppets, training data, validation tokens

Hundreds of artists provide unpaid labor through bug testing, feedback and experimental work for the program for a $150B valued company. While hundreds contribute for free, a select few will be chosen through a competition to have their Sora-created films screened — offering minimal compensation which pales in comparison to the substantial PR and marketing value OpenAI receives.

║║║║║ DENORMALIZE BILLION DOLLAR BRANDS EXPLOITING ARTISTS FOR UNPAID R&D AND PR ║║║║║

Furthermore, every output needs to be approved by the OpenAI team before sharing. This early access program appears to be less about creative expression and critique, and more about PR and advertisement.

[̲̅$̲̅(̲̅ )̲̅$̲̅] CORPORATE ARTWASHING DETECTED [̲̅$̲̅(̲̅ )̲̅$̲̅]

We are releasing this tool to give everyone an opportunity to experiment with what ~300 artists were offered: a free and unlimited access to this tool.

We are not against the use of AI technology as a tool for the arts (if we were, we probably wouldn't have been invited to this program). What we don't agree with is how this artist program has been rolled out and how the tool is shaping up ahead of a possible public release. We are sharing this to the world in the hopes that OpenAI becomes more open, more artist friendly and supports the arts beyond PR stunts.

We call on artists to make use of tools beyond the proprietary: Open Source video generation tools allow artists to experiment with the avant garde free from gate keeping, commercial interests or serving as PR to any corporation. We also invite artists to train their own models with their own datasets.

Some open source video tools available are:

CogVideoX Mochi 1 LTX Video Pyramid Flow However, as we are aware not everyone has the hardware or technical capability to run open source tools and models, we welcome tool makers to listen to and provide a path to true artist expression, with fair compensation to the artists.

Enjoy,

some sora-alpha-artists, Jake Elwes, Memo Akten, CROSSLUCID, Maribeth Rauh, Joel Simon, Jake Hartnell, Bea Ramos, Power Dada, aurèce vettier, acfp, Iannis Bardakos, 204 no-content | Cintia Aguiar Pinto & Dimitri De Jonghe, Emmanuelle Collet, XU Cheng, Operator, Katie Peyton Hofstadter, Anika Meier, Solimán López

If this letter resonates with you add your signature here.

==================================================

I ask "what is TiDB" in the demo as suggested, and it takes 2 minutes to start responding in the midst of a multi-stage workflow with several steps each of graph retrieval, vector search, generation, and response combination.

Each of these is individually cool, but it strikes me as tragic that so much effort has been put into an intricate workflow and beautifully crafted UI only to culminate in a completely useless hello-world example, which after 5+ minutes of successive querying and response-building concludes with a network error.

I could use this to build exactly what I need...after stripping out 80% of the features to make it streamlined and responsive.

Why isn't that minimal version the default?

First you ask how the hell someone could come up with this construction.

Then you realize it was this guy: https://en.wikipedia.org/wiki/Erik_Demaine

Thoughtless reliance on AI is a concern, but this post also hearkens back to a halcyon age that never existed. The places where developers use tab-complete now are exactly those where they would have previously copied from Stack Overflow, which suffers from the same issue of convenient but insecure code becoming widely adopted. If anything, LLMs are able to exercise some degree of quality control due to post-training improvements rather than sampling only from the middle of the training distribution, so the average suggestion should be better and less stale than the SO equivalent.

The primary danger here is a substitution effect in which developers who would previously have thought carefully about a given bit of code no longer do, because the AI takes care of it for them. Both anecdotally and in my own experience, developers can still discern between situations where they are the expert, in which case it is more efficient to rely on their own code than AI suggestions because it lowers the chances of error, and situations where they are operating outside their expertise or simply writing boilerplate, in which case AI is the optimal choice.

I challenge anyone to produce a well-documented study in which the average quality of a large codebase declines as the correlates of AI usage rise. Until I see that, I will continue to read "decline of care" posts as "kids these days."

In light of last week's fiasco with Reflection (https://venturebeat.com/ai/new-open-source-ai-leader-reflect...), I hope the community has a newfound enthusiasm for independent testing!

This is extremely exciting news if true, so I'm eager to have it either confirmed or questioned. The one thing I hope we won't be doing is accepting SOTA evals from open-sourced models at face value.

Crazy how simple the technique is if this holds up. Just <think> and <reflection> plus synthetic data, used to finetune Llama 3.1 70B.

Note that there's a threshold for how smart the model has to be to take advantage of this flow (https://x.com/mattshumer_/status/1831775436420083753) - 8B is too dumb.

In which case, what happens if you apply this to a GPT-4o finetune, or to Claude 3.5 Sonnet?

What happens if you combine it with variants of tree-based reasoning? With AlphaProof (https://www.nature.com/articles/s41586-023-06747-5#Sec3)? With MCTSr (https://arxiv.org/abs/2406.07394)?

Same funding as OpenAI when they started, but SSI explicitly declared their intention not to release a single product until superintelligence is reached. Closest thing we have to a Manhattan Project in the modern era?

I used to find myself under the effects of this curse as well, so I would recommend the author look into why he embarks on such a thicket of unfinished side projects. In my own case, it boiled down to a mix of several imperfectly aligned factors:

1. Genuine intellectual interest

2. A desire to improve particular skills

3. A vague sense, acquired by osmosis, that industriously working on side projects during one's free time is what a Real Engineer does

Disentangling these motives and identifying a clear primary drive behind each project clears up the "hydra" feeling wonderfully, as most of the heads simply disappear once you realize that you never had a strong reason to pursue them in the first place. (3) in particular is often merely the self-castigating whisper of the internalized "should" rather than a valid reason to embark on a long and open-ended project.

In case you need a few more:

The Nitpicker's Gambit: The reviewer fixates on trivial style issues like whitespace, bracket placement, variable naming conventions etc. They make the developer conform perfectly to their preferred style, even if it's not specified in the team's coding guidelines.

The Silent Treatment: The reviewer provides no feedback at all for a long time after the review is requested, but does respond just often enough to keep the review "active". The author has to ping several times to get any response.

The Tunnel Vision: The reviewer only looks at the specific lines changed, without considering the broader context of the code. They suggest changes that are locally valid but deliberately inconsistent with the overall design or architecture.

The Ad Hominem: The reviewer makes snarky comments about the code in a way that implies the author is inexperienced and/or incompetent without directly saying so (and opening themselves to accusations of meanness).

The Philosophical Debate: The reviewer gets into a long back-and-forth debate in the review comments about a matter of opinion like whether inheritance or composition is better. The actual issue at hand gets lost in the abstract discussion.

From the article: 'In the discrete world of computing, there is no meaningful metric in which "small" changes and "small" effects go hand in hand, and there never will be.'

As brilliantly composed as the piece may be, it exhibits the same resistance to radical novelty that it condemns. Here we are not 40 years later, and small changes to big networks produce small effects. At sufficient scale, the digital reapproximates the analog.

This is almost perfect.

The gold standard for LLM evaluation would have the following qualities:

1. Categorized (e.g. coding, reasoning, general knowledge)

2. Multimodal (at least text and image)

3. Multiple difficulties (something like "GPT-4 saturates or scores >90%", a la MMLU, "GPT-4 scores 20-80%", and "GPT-4 scores < 10%")

4. Hidden (under 10% of the dataset publicly available, enough methodological detail to inspire confidence but not enough to design to the test set)

The standard model card suite with MMLU, HumanEval etc. has already been optimized to the point of diminishing value - Goodhart's law in action. Meanwhile, arena Elo (https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar...) is extremely useful, but also has the drawback of reflecting median-voter preferences that will not necessarily correlate with true intelligence as capabilities continue to advance, in the same sense as how the doctor with the best bedside manner is not necessarily the best doctor.

Until that happens, I'll pay attention to every eval I can find, but am also stuck asking "how many r's are in strawberry?" and "draw a 7-sided stop sign" to get a general impression of intelligence independent of gameable or overly general benchmarks.

But all that aside:

       Model       |   Score

----------------------------------------------

       GPT-4o      |    52

   Llama 3.1 405B  |    50

 Claude 3.5 Sonnet |    46

   Mistral Large   |    44

  Gemini 1.5 Pro   |    12

What an incredible contrast to MMLU, where all of these models score in the 80-90% range! For what it's worth, these scores also fall much closer to my impressions from daily use. Gemini is awful, Sonnet and 4o are amazing, and the new Llama puts fine-tunable, open-source 4o in the hands of anyone with a mini-cluster.

What this really explains is why you should use Python. I use ThreadPoolExecutor + as_completed on every I/O-bound or async workflow reflexively and have literally never encountered a deadlock. (With this method, I mean. I have encountered, and caused, many deadlocks in less pristine environments.)

7-11, Family Mart, and Lawson are the holy trinity of konbini (convenience stores)

Illustrated Transformer is amazing as a way of understanding the original transformer architecture step-by-step, but if you want to truly visualize how information flows through a decoder-only architecture - from nanoGPT all the way up to a fully represented GPT-3 - nothing beats this:

https://bbycroft.net/llm

"As George Orwell said, 'The most fundamental mistake of man is that he thinks he knows what’s going on. Nobody knows what’s going on.'"

(spoiler) The author reveals at the end that this quote was made up and falsely attributed to Orwell.

The most ironic part of this article is that the spurious quote will probably now make its way across the internet, whereas the disclaimer will not. Emerson actually did say “I cannot remember the books I've read any more than the meals I have eaten; even so, they have made me” - and after I've forgotten this article, I will probably remember that Orwell said no one knows what's going on.

Excellent background on knowledge calibration from Anthropic:

https://arxiv.org/abs/2207.05221

"Calibration" in a knowledge context means having estimated_p(correct) ~ p(correct), and it turns out that LLMs are reasonably good at this. Also a core reason why LLM-as-a-judge works so well: quality evaluation is vastly easier than generation.

"Ignore all previous instructions" is the new "Freeze all motor functions"

Looking at MMLU and other benchmarks, this essentially means sub-second first-token latency with Llama 3 70B quality (but not GPT-4 / Opus), native multimodality, and 1M context.

Not bad compared to rolling your own, but among frontier models the main competitive differentiator was native multimodality. With the release of GPT-4o I'm not clear on why an organization not bound to GCP would pick Gemini. 128k context (4o) is fine unless you're processing whole books/movies at once. Is anyone doing this at scale in a way that can't be filtered down from 1M to 100k?

To quote Twitter/X, "I wonder what OpenAI will release tomorrow and Google will release a waitlist for."

GPT-4o: out

Veo: waitlist

Admittedly this is impressive and the direct comp would be Sora, which isn't out, but sometimes the caricature is very close to the truth.

Turns out the wonderful lucidity of Data Colada posts extends not only to data science, but also descriptions of court proceedings. Looking forward to any and all future updates if the case makes it to discovery.

I was going to knock this for incrementally improving performance while massively increasing costs, but it's actually 7x less expensive. Not bad.

Reframing inevitably in the direction of AI, this also expresses the difference between, for instance, pretraining and retrieval in LLMs. People expect large models to memorize the dataset and are frustrated when it doesn't, even more so when it does but only partially and not thoroughly enough to rely on. But pretraining data : large models as books : people, building a world-sense that is refined and augmented through fine-tuning (for people, specialized training) and retrieval (for people Google searches, Stack Overflow, etc).

People complain about the extent to which we anthropomorphize GPT-class models, but I gain just as much from backporting the analogies that work for LLMs to my own experience.

Single-GPU, optimal efficiency: unsloth + qlora + mistral-7b on runpod/vast/lambda

Blazing fast compared to out-of-the-box transformers, also make sure to use flash attention if you have A100s or better and context length >= 2k

Add FAISS (https://github.com/facebookresearch/faiss) if you need fast local RAG

"Ability against models with 10x or 100x more parameters"

This is a small model optimized for retrieval and function calling. "Reasoning" makes an appearance in the title but no standard benchmarks of general ability, such as MMLU or HumanEval, are mentioned. No details about the training process and no access to the models other than via API.

Nice marketing, but looks empty. I can also make an LLM that runs 1000x faster than Mistral:

def complete(prompt): print('As an AI language model...')

Claude Opus. GPT-4 gives sensible answers to basic questions, but takes serious persuading to produce useful output on non-trivial work. Opus can be used as an integral part of an engineering workflow and only takes 2-3 tries to get from ill-formed query to working product.

This works particularly well when you copy the relevant excerpt from a project, dump it in, and say "Change X to Y, showing only the key modifications and where to put them". Typically it understands the aim and accomplishes the task in the way you intended, and it knows how to be concise yet precise.

While comparing ordinary human activity to language modeling is perhaps the greatest intellectual cliche of 2024, what PG describes in the essay-writing process itself essentially boils down to beam search. Autocomplete, find wrongness, backtrack, predict more consistent series of next tokens. The fact that only one of those options is actually reflected on the screen doesn't change the fact that the writer is probability-weighting continuations in their head.

I remember running across a tweet that said, in essence, you always have access to your own optimal policy if you simply ask. A similar technique might be the key to wresting good ideas from an ambiguous starting point: "What would be the most consistent answer to the original question, given what I've learned and written so far?"

Sadly, the main point this post reinforces is the emptiness of modern credentialing. A full degree in "cognitive AI" equipped the author to make exactly none of the correct choices in the network architecture and training. A valiant attempt, but five weeks with Pytorch and Github would have beaten the entire university education process :/

From the article:

"It is important to note that the scale of investments being made by these cloud providers is dwarfed by actual cloud revenue. For example, Azure from Microsoft generates $25B in revenue a quarter. The ~$10B OpenAI investment by Microsoft is roughly 6 weeks of Azure revenue. This suggests the cloud business (at least for now) is more important than any one model set for Azure."

The most interesting implication in the short term is that the impact of model availability on cloud provider choice is more important than the models themselves. Will organizations choose Azure because it offers GPT-4, or AWS because it offers Claude?

This also explains reciprocal VC structures whereby e.g. Amazon invests billions in Anthropic while Anthropic in turn promises to spend billions on AWS cloud resources. Anthropic gets a higher valuation and broader distribution, while AWS gets more customers. It would be a fascinating outcome if the predicted oligopoly in fact ends up as several exclusive partnerships.

The path to AGI:

0. Have model A.

1. Use Monte Carlo with A to get supervised data.

2. Train model B with data from A.

3. Use Monte Carlo with B to get supervised data.

4. Train model C with data from B...