Dmitriy Karpman

Ilya Sutskever confirmed at NeurIPS '24 pre-training scaling law will not be the same. Internet data, the AI fossil fuel, is running out. It is not the end of the AI hype, the rapid rise of foundational AI has outpaced our ability to fully grasp their long-term societal impact. AI agents and applications are poised to transform every facet of our lives, and we will see it happening in coming few years. There will still be improvements in model architecture, and synthetic data creation, so models will improve, just not at the lightning speed they did till now. Looking forward, the single most important milestone in AI is creating systems capable of advancing themselves—AI researchers that can innovate and improve foundational models autonomously. We need to assess how far are we from this stage.

9

Today, I found a fascinating way to think about “Tokens” in LLMs. It came from a conversation between Jensen Huang and Patrick Collison at a recent Stripe event. When asked about the future of compute capacity in five years, Jensen smartly dodged that question a bit, but then gave an excellent explanation and analogy about why tokens are the new forces that will power the next decades of humanity. He explained that we're now producing something unprecedented (and at scale): floating point numbers that possess “value”, which we now call tokens. These tokens are valuable because they encapsulate “intelligence”. People are now taking these tokens and transforming them into English, French, images, videos, chemicals, proteins, robotic movements, etc. And many are working hard to expand the range of concepts and ideas we can create with these tokens. He then goes on to made a compelling comparison between tokens and electricity. In the past industrial revolution, we successfully found a way to convert “atoms” into “electrons” (by boiling water to power electricity turbines). And now, we've discovered a way to convert "electrons" into "tokens" (by using energy to power data centers that train and run LLMs). When electricity was first introduced, few people understood its value. Today, paying for kilowatts is routine. The same will happen with tokens. Right now, only early adopters and builders are paying tokens. Soon, everyone will be paying for tokens on a daily basis to supercharge productivity and power new products and services. Many new industries will be born from and built on top of tokens. When I first heard about this way of thinking, I got goosebumps. Perhaps it’s because Jensen has a good storytelling skill that helps him sell what he builds, but I have to admit this way of thinking gave me a brief surge of pride. I'm proud of humanity's collective effort and creativity, which have taken us from living under rocks to building machines that can "think." That's nothing short but a miracle.

18

5 Comments

🎯Data was, is, and will be the only fuel. Analogy with the fossil fuel is a good one, since the data we have today is as dirty as oil and coal. What big tech companies tried is to put a lot of "garbage" data, and what they did get was fantastic. Yet, we are at the point of diminishing returns. Data is the beginning and the end of this cycle.” Great thread, particularly this comment 👆from Ilija Lazarevic

20

2 Comments

Here's a project I did this weekend analyzing the type of state-enforced bias and censorship on Chinese LLMs. This is probably going to be more important moving forward as more of these are not only quite strong (Qwen2 7B Instruct, which I tested is competitive with Llama 3 8B) but more importantly, being liberally licensed - Yi 1.5 and Qwen2 are both largely Apache 2.0 licensed. These are also multi-lingual models, which means it's probably going to end up being used for all kinds of things without downstream users' knowledge. Anyway, in for a penny, in for a pound, so figure I might as well share this widely, since I haven't seen a similar writeup/exploration on this yet and IMO the results are quite interesting: https://lnkd.in/g_WKscA6 Besides the writeup I also applied the recent "abliteration" (orthogonalized refusal vector removal) technique to create a model that largely (~80%) removed refusals without significantly affecting performance and a "deccp" dataset as a starting point for those looking to play around with testing Chinese state censorship alignment of their models. Model, dataset, and all related code are Apache 2.0 licensed and can be found linked from the article.

17

3 Comments

Cost-aware inference is important, but currently LLM prompts are not interoperable. Prompts that work good on one LLM may show worse performance on the other, even more advanced LLM or even on the different versions of the same LLM. Once interoperability problem is solved it will also be possible to implement cost-aware inference that strategically manages computational resources for prompt processing, ensuring sustainable costs. This approach includes model routing https://www.notdiamond.ai/ - automatically selecting the most appropriate model for each task. Startups in this area are actively supported by the industry veterans.

4

2 Comments

DeepSeek V3 has achieved state of the art performance at a fraction of the cost of other frontier models. It cost them $5.6M in compute time to train from scratch. GPT 3.5 cost $250M and was just released 2 years ago. Unreal.

11

3 Comments

You don’t normally see LLM researchers and engineers talking about their day-to-day work. Big shoutout to this YouTube video from the Anthropic LLM Interpretability team, the team behind the Golden Gate Claude. It’s amazing to see that most of the team members have worked for the team for only one year, and yet they have made such a big breakthrough in interpretability research. Another thing I want to highlight is the unique challenges facing the team and how they approach these challenges, that is, through fast trial-and-error iterations and constant evaluation of what may work and what to prioritize next. #artificialintelligence #machinelearning #deeplearning https://lnkd.in/eUaPAYch

97

1 Comment

If you thought software engineers were grumpy before, just wait until they start working with open-ended AI systems. One of the biggest challenges in designing AI systems is establishing guardrails to the behaviors. For example, in traditional software, you can click one of two buttons...two options developers have to deal with. In language-based systems, the inputs are theoretically infinite, which means I have to try to constrain the user to a semi-finite set of options. Without limits, users get frustrated by bad responses, hallucinations, and unintended consequences. In short, engineers now need to enforce policies instead of explicit rules and that is a substantial increase in complexity (and therefore grumpiness). https://buff.ly/3KvNOnq

3

🚀 Unlocking the Power of Variational Autoencoders in Time Series Analysis! 📊 I'm excited to share my latest Substack post, where I dive deep into the fascinating world of Variational Autoencoders (VAEs) and their application in generating synthetic time series data. 🔍 What You'll Discover: A clear explanation of how VAEs work, including the encoder and decoder architecture. Step-by-step code to implement a VAE for time series data generation, complete with visualizations of the results! This post is perfect for data scientists, machine learning enthusiasts, or anyone curious about generative models and their practical applications. Let’s unlock new insights in data generation together! 🌟 #MachineLearning #DataScience #VariationalAutoencoder #TimeSeries #AI #DeepLearning #DataGeneration

2

🚨🚨 🚨 Unlocking Innovation: The Power of the Adjacent Possible Innovation programs can stall when the team tries to ideate way beyond their collective perspective and frame of reference. Like a software engineer brainstorming genetic modification of mycelia, or a senior executive ideating process improvements which they haven't been participated in for decades. It's challenging, exhausting, and ultimately self-defeating. Enter the concept of the "Adjacent Possible" - a game-changer for innovation. What is it? The Adjacent Possible, coined by Stuart Kauffman, represents the realm of new ideas within reach based on our current knowledge, resources, and capabilities. Why it matters: ➡️ Realistic Innovation: it focuses on achievable advancements, not far-fetched concepts. ➡️ Incremental Progress: each innovation opens doors to new possibilities. How to leverage it: • Constantly explore the boundaries of your adjacent possible. What's in your line of sight is what you know best. • Combine existing ideas in novel ways. Even small perturbations from status quo can generate large innovation leaps. • Stay curious about developments in adjacent industries. Cross-domain patterns create fresh sources of ideas, while also expanding the adjacent possible. Most of all, it takes the 'genius' pressure off the team. No one knows everything, but each one of us knows a lot about something, and we can all drive plenty of innovation potential by focusing on the spaces we know best. Are you exploring your Adjacent Possible? Share your thoughts below! 👇 #Innovation #innovationsystems #AdjacentPossible

8

2 Comments

My third essay for Yule LOGIC is about LLMs and temperature: https://lnkd.in/gqutMN-R

10

1 Comment

We are quickly entering the "trough of disillusionment" in the hype cycle for generative AI. While GPTs excel in automating certain tasks (and can even automate away large portions of some jobs) they often fall short in others. Like all new technology, the issue lies in misapplying generative AI for tasks it wasn't designed for. This is a problem compounded by the GPTs themselves suggesting they can handle these tasks (often cheerfully!). This misalignment creates frustration and unmet expectations which we're starting to see more of these days. As the dust settles I firmly believe we'll recognize GPTs as a transformative technology across many industries. However it's crucial to understand their limitations and apply them appropriately. What are your thoughts on the current state of generative AI and its future applications?

2

Hugging Face demonstrated test-time scaling with Llama 1B, turning it into a "reasoner" that outperformed Llama 8B (8x larger!) in math. Test-time scaling has opened up a whole new dimension in model performance and AI research. I wrote about this when OpenAI released its o-1 model (the first 'reasoner' model). You can read it here: https://lnkd.in/gi737nsA At its core, a reasoner model uses Chain of Thought (CoT) reasoning—breaking problems into logical steps, solving them sequentially, and backtracking when needed. Scaling test-time compute enables this reasoning through two techniques: 1/ Self-Refinement: Models iteratively improve outputs, correcting errors over multiple passes. 2/ Search & Verifiers: Generate multiple outputs and evaluate them with verifier models (e.g., Process Reward Models for steps, Outcome Reward Models for final answers). Hugging Face’s contribution stands out because they’ve shown test-time scaling can be applied to any model, even smaller, open-source ones. They’ve also made it easier for others to replicate and advance the technique—helping the OS community catch up. That said, inference scaling is still far from efficient. Llama 1B required ~20 reasoning steps to outperform Llama 8B, a model only 8x larger. Nevertheless, inference scaling already achieves two critical things: – Pushing the frontier: Scaling inference pushes the boundaries of accuracy and delivers better results. – Smaller hardware requirement: Smaller models can match the performance of larger ones, reducing reliance on expensive infrastructure. The next frontier isn’t just about bigger models—it's about smarter inference. Excited to see where this leads!

62

4 Comments

One new interesting application of LLMs: CAIS has released a forecasting bot that provides probability estimates for any question you ask. For example, to the question "Will GPT5 be released before 2025?", it answers that it's 70% likely. The bot works by searching for relevant information online, feeding it to GPT-4o, asking a few guiding questions, and then requesting a probability estimate. What's impressive is that this bot outperforms experienced human forecasters, even though GPT-4o wasn't specifically trained for this task. This kind of application shows that AI could be of great help in improving decision-making in the future. Link to the blog post: https://lnkd.in/gjP7JQgz Link the bot if you want to try it yourself: https://forecast.safe.ai/

11

After Apple’s research saying LLMs can’t think, it’s funny that now all I’m seeing is research on teaching them to think – something most humans still struggle with. Papers: Looking Inward: Language Models Can Learn About Themselves by Introspection https://lnkd.in/gf6rgnUP On The Planning Abilities of OpenAI's o1 Models: Feasibility, Optimality, and Generalisability https://lnkd.in/gNsPtDbi Thinking LLMs: General Instruction Following with Thought Generation https://lnkd.in/guQtRnsD #AIResearch #MachineLearning #ArticificalIntelligence

24