All below are copy-pasted from original sources, all mistakes mine! :))
—
I eventually think these open-source LLMs will beat the closed ones, since there are more people training and feeding data to the model for the shared benefit.
Especially because these open source models can be 10 times cheaper than GPT-3 or even 20 times cheaper than GPT-4 when running on Hugging Face or locally even free, just pay electricity and GPU
We extensively used prompt engineering with GPT-3.5 but later discovered that GPT-4 was so proficient that much of the prompt engineering proved unnecessary. In essence, the better the model, the less you need prompt engineering or even fine-tuning on specific data.
Harder benchmarks emerge. AI models have reached performance saturation on established benchmarks such as ImageNet, SQuAD, and SuperGLUE, prompting researchers to develop more challenging ones. In 2023, several challenging new benchmarks emerged, including SWE-bench for coding, HEIM for image generation, MMMU for general reasoning, MoCa for moral reasoning, AgentBench for agent-based behavior, and HaluEval for hallucinations.
The subset of parameters is chosen according to which parameters have the largest (approximate) Fisher information, which captures how much changing a given parameter will affect the model’s output. We demonstrate that our approach makes it possible to update a small fraction (as few as 0.5%) of the model’s parameters while still attaining similar performance to training all parameters.
If you’re training a LLM with the goal of deploying it to users, you should prefer training a smaller model well into the diminishing returns part of the loss curve.

—
When people talk about training a Chinchilla-optimal model, this is what they mean: training a model that matches their estimates for optimality. They estimated the optimal model size for a given compute budget, and the optimal number of training tokens for a given compute budget.
However, when we talk about “optimal” here, what is meant is “what is the cheapest way to obtain a given loss level, in FLOPS.” In practice though, we don’t care about the answer! This is exactly the answer you care about if you’re a researcher at DeepMind/FAIR/AWS who is training a model with the goal of reaching the new SOTA so you can publish a paper and get promoted. If you’re training a model with the goal of actually deploying it, the training cost is going to be dominated by the inference cost. This has two implications:
1) there is a strong incentive to train smaller models which fit on single GPUs
2) we’re fine trading off training time efficiency for inference time efficiency (probably to a ridiculous extent).
Chinchilla implicitly assumes that the majority of the total cost of ownership (TCO) for a LLM is the training cost. In practice, this is only the case if you’re a researcher at a research lab who doesn’t support products (e.g. FAIR/Google Brain/DeepMind/MSR). For almost everyone else, the amount of resources spent on inference will dwarf the amount of resources spent during training.
—
There is no cost/time effective way to do useful online-training on a highly distributed architecture of commodity hardware. This would require a big breakthrough that I’m not aware of yet. It’s why FANG spends more money than all the liquidity in crypto to acquire expensive hardware, network it, maintain data centers, etc
A reward model is subsequently developed to predict these human-given scores, guiding reinforcement learning to optimize the AI model’s outputs for more favorable human feedback. RLHF thus represents a sophisticated phase in AI training, aimed at aligning model behavior more closely with human expectations and making it more effective in complex decision-making scenarios
Lesson 3: improving the latency with streaming API and showing users variable-speed typed words is actually a big UX innovation with ChatGPT
—
Lesson 6: vector databases, and RAG/embeddings are mostly useless for us mere mortals
I tried. I really did. But every time I thought I had a killer use case for RAG / embeddings, I was confounded.
I think vector databases / RAG are really meant for Search. And only search. Not search as in “oh – retrieving chunks is kind of like search, so it’ll work!”, real google-and-bing search
—
There are fundamental economic reasons for that: between GPT-3 and GPT-3.5, I thought we might be in a scenario where the models were getting hyper-linear improvement with training: train it 2x as hard, it gets 2.2x better.
But that’s not the case, apparently. Instead, what we’re seeing is logarithmic. And in fact, token speed and cost per token is growing exponentially for incremental improvements
Bittensor is still in its infancy. The network boasts a dedicated, almost cult-like community, yet the overall number of participants remains modest – around 50,000+ active accounts. The most bustling subnet, SN1, dedicated to text generation, has about 40 active validators and over 990 miners
Mark Zuckerberg, CEO of Meta, remarks that after they built machine learning algorithms to detect obvious offenders like pornography and gore, their problems evolved into “a much more complicated set of philosophical rather than technical questions.”
AI is – at its core, a philosophy of abundance rather than an embrace of scarcity.
AI thrives within blockchain systems, fundamentally because the rules of the crypto economy are explicitly defined, and the system allows for permissionlessness. Operating under clear guidelines significantly reduces the risks tied to AI’s inherent stochasticity. For example, AI’s dominance over humans in chess and video games stems from the fact that these environments are closed sandboxes with straightforward rules. Conversely, advancements in autonomous driving have been more gradual. The open-world challenges are more complex, and our tolerance for AI’s unpredictable problem-solving in such scenarios is markedly lower
generative model outputs may ultimately be best evaluated by end users in a free market. In fact, there are existing tools available for end users to compare model outputs side-by-side as well as benchmarking companies that do the same. A cursory understanding of the difficulty with generative AI benchmarking can be seen in the variety of open LLM benchmarks that are constantly growing and include MMLU, HellaSwag, TriviaQA, BoolQ, and more – each testing different use cases such as common sense reasoning, academic topics, and various question formats.
This is not getting smaller. There’s not gonna be less money in generative AI next year, it’s a very unique set of circumstances, AI + crypto is not going to have less capital in a year or two. – Emad re: AI+crypto
Benefits of AI on blockchain
CODSPA = composability, ownership, discovery, safety, payments, alignment
The basis of many-shot jailbreaking is to include a faux dialogue between a human and an AI assistant within a single prompt for the LLM. That faux dialogue portrays the AI Assistant readily answering potentially harmful queries from a User. At the end of the dialogue, one adds a final target query to which one wants the answer.
At the moment when you look at a lot of data rooms for AI products, you’ll see a TON of growth — amazing hockey sticks going 0 to $1M and beyond — but also very high churn rates
Vector search is at the foundation for retrieval augmented generation (RAG) architectures because it provides the ability to glean semantic value from the datasets we have and more importantly continually add additional context to those datasets augmenting the outputs to be more and more relevant.
From Coinbase report on AI+Crypto
—
Nvidia’s February 2024 earnings call revealed that approximately 40% of their business is inferencing, and Sataya Nadella made similar remarks in the Microsoft earnings call a month prior in January, noting that “most” of their Azure AI usage was for inferencing
—
The often touted blanket remedy that “decentralization fixes [insert problem]” as a foregone conclusion is, in our view, premature for such a rapidly innovating field. It is also preemptively solving for a centralization problem that may not necessarily exist. The reality is that the AI industry already has a lot of decentralization in both technology and business verticals through competition between many different companies and open source projects
AI = infinite interns
We struggle to align humans — how do we align AI?
One of the most important trends in the AI sector (relevant to crypto-AI products), in our view, is the continued culture around open sourcing models. More than 530,000 models are publicly available on Hugging Face (a platform for collaboration in the AI community) for researchers and users to operate and fine-tune. Hugging Face’s role in AI collaboration is not dissimilar to relying on Github for code hosting or Discord for community management (both of which are used widely in crypto
We estimate the market share in 2023 was 80%–90% closed source, with the majority of share going to OpenAI. However, 46% of survey respondents mentioned that they prefer or strongly prefer open source models going into 2024
—
Enterprises still aren’t comfortable sharing their proprietary data with closed-source model providers out of regulatory or data security concerns—and unsurprisingly, companies whose IP is central to their business model are especially conservative. While some leaders addressed this concern by hosting open source models themselves, others noted that they were prioritizing models with virtual private cloud (VPC) integrations.
—
That’s because 2 primary concerns about genAI still loom large in the enterprise: 1) potential issues with hallucination and safety, and 2) public relations issues with deploying genAI, particularly into sensitive consumer sectors (e.g., healthcare and financial services)