Simon Suo

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

Some takeaways and opportunities for new AI product tooling from my time at KubeCon last week: 🖥️ CPU Inference and On-Prem Deployments: While we're still in the early innings of GPU Inference optimization, there are opportunities worth building towards in the CPU inference space as well. Optimizing CPU inference on-premises is critical for reducing cloud reliance, improving data privacy, and cutting costs. Companies aim to streamline local compute efficiency by enabling faster, more resilient operations with reduced network dependency. 🎮 Texture Generation and Modeling Behavior in AI Systems: Advanced AI techniques, such as texture and mixed-generative models, are proving useful in applications requiring nuanced behavioral analysis and are particularly valuable in industries like gaming and simulation, where creating realistic environments and interactions is crucial. However, training these AI models often requires vast datasets and considerable processing power. Infrastructure teams must ensure that these models can operate at scale, often requiring hybrid solutions that span both on-prem and cloud environments. 👀 Observability and Real-Time Monitoring on multi-format data: AI-powered observability tools can aggregate and analyze systems across data from kernel, applications, and infrastructure layers and do so in real-time, improving troubleshooting and lowering MTTR (mean time to response). However, balancing insights with rising data storage costs remains a key challenge for enterprises. 💻 Modernizing Codebases with AI Transformation Tools: AI tools are helping modernize legacy codebases, reducing technical debt and easing maintenance. While adoption challenges persist, these tools are evolving to make code updates faster and more efficient. Solutions like Moderne, and Codemod are enabling engineers to translate outdated code into modern, efficient formats with minimal manual intervention. 💡 AI-Driven Cost Management and Efficiency: Infrastructure costs are continuing to skyrocket as companies scale up, particularly in terms of data storage, compute, and network usage. AI offers promising opportunities to optimize these expenses through predictive cost management, dynamic resource allocation, and automated scaling. By implementing tools that forecast usage trends and recommend cost-saving measures, such as with CloudZero organizations can gain a clearer picture of their resource consumption. This proactive stance on cost management allows for more efficient resource utilization, ensuring that infrastructure spending aligns with actual demand. AI is no longer a nice-to-have in the infrastructure software space; it’s quickly becoming a competitive necessity. If you're a founder building in any of these areas or themes, I'd love to chat.

48

3 Comments

AI memory is fundamental for conversational AI and AI agents. It is critical for building personalized experiences and improving reasoning. Check out Letta’s course to understand how to use AI memory to enhance the customization and quality of your AI app.

15

🎯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

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

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

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

I’ve been fortunate enough to do AI shows with Sam Altman, Yann LeCun & Arthur Mensch. But this show today with David Cahn at Sequoia Capital is the best of all of them: 🧠 AI’s $600BN question 🏃 Why no one will never train a frontier model on the same data center twice 🏗️ Servers, steel & power: The 3 pillars of the future of AI Not one to be missed, my 8 key takeaways 👇 1. The Oligopoly That is Blowing Billions on AI - Microsoft, Azure, & Google now represent $7T market cap. - Of course they will spend billions to protect their oligopoly - What is most interesting is Meta is the only incumbent that is financing this not with cloud being their cash cow. 2. How Startups Could Win with the Cost of Compute Coming Down - Big tech companies are producers of compute & startups are consumers of compute. - Overproduction of compute = lower prices.  - This gives startups higher margins & makes them more valuable. 3. Is Compute the Currency of the Future or Do the Costs Come Down and It Is Commoditised? - We are forgetting “compute” is a euphemism for a $2BN data center in the middle of nowhere. - They have GPUs & liquid cooling systems that need to be constantly optimized & upgraded. - We cannot simply build 15 years worth of compute all at once.  No one's ever going to train a frontier model on the same data center twice. 4. Data Centre Teams and Model Teams Need to Be Coupled - You cannot have separate teams running data centers & building models. - Elon & Zuck control all their data centres. - It will not work as these models get better & better. 5. Why the Data Center Is the Most Important Asset - There is not much difference between these models today. - AI researchers are jumping from one lab to another. - Scaling laws will become more dominant as these models get bigger. 6. Servers Steel and Power: The Three Bottlenecks of AI Three things: 1. Servers: NVIDIA, AMD, Broadcom… 2. Steel: Construction & real estate developers for data centers. 3. Power: All the energy required for AI. - More factories will be built in 12 months for AI and we will see another industrial revolution. 7. You Have to Have a Cash Machine to Play the Big Model Game - Amazon has AWS - Meta has Instagram - Microsoft has Azure - You need a cash machine to compete and it CANNOT be the AI business. 8. There is One Definition of Success in Venture: Are you a Slugger? - If you are generating billion dollar gains, you are a slugger. - Doug Leone, Pat Grady, Andrew Reed are all sluggers. - You cannot be good at this business without being a slugger. (links in comments) #founder #funding #business #investing #vc #venturecapital #entrepreneur #startup #seed #funding #sequoia

190

26 Comments

🚨🚨 🚨 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

NVIDIA’s latest LLM, Hymba isn’t just another LLM. It’s rewriting the playbook for small models A hybrid Mamba x Attention architecture with only 1.5 billion parameters but punches above its weight, outperforming larger small LLMs like Meta 3.2 and SmolLM v2.  The twist? It’s trained on just 1.5T tokens.  Here’s why that’s exciting:  - Parallel hybrid heads: Combines Mamba and Attention heads to enhance performance.    - Meta tokens: Prepended learnable tokens fine-tune the model dynamically.  - Efficient KV cache sharing: Across layers and heads for faster context handling.  It also includes:  - 16 SSM states  - 3 full attention layers (others use sliding window attention).  - FlexAttention support from PyTorch 2.5 for better adaptability.  Weights are coming soon on Hugging Face.  What’s better? It's commercially usable.  Full details + repo here: https://lnkd.in/e47xCPcS  ___________  I share my learning journey here. Join me and let's grow together.  Enjoy this? Repost it to your network and follow Karn Singh for more.  

14

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

LLMs are terrible at basic math. Why? Here's why: Unlike people, LLMs don’t conceptualize mathematics. They simply recognize patterns in the training data (text) they’ve learned from. When asked what 2 + 2 is, an LLM recalls instances of seeing ‘2’, ‘+’, ‘2’ and so on, allowing it to approximate solutions but not match the precision of a calculator. That’s how we end up with LLMs asserting that 2 + 2 = 5. But for analytical teams, like finance, that makes LLMs really hard to adopt in day-to-day workflows. The productivity gains of having AI do calculations are negated by needing to fact-check and audit every calculation and number. LLMs are also not deterministic, so no amount of prompting is going to guarantee you’ll get the correct answer every time. And that’s just for basic math – the outputs I’ve seen from LLMs on more complex calculations and analyses can get very wonky. Add thousands of rows of data and you’re squarely in hallucination land. So what’s the solution? Calculate things manually? No. The good news is that this is changing quickly and there are options: 1 - AI is getting better at math. Researchers continue to train LLMs with more specialized datasets that include a much wider array of mathematical problems and solutions. The newest models, like OpenAI’s latest family of models, are incorporating things like symbolic reasoning and “chain of thought” methods to break complex problems into smaller, logical steps the LLM can more easily process. Results have been really promising and we’ve seen each new iteration of models continue to beat previous benchmarks. 2 - Leverage AI agents For intensive workflows, and to really make sure your outputs are correct, working with companies that are building high-performing AI agents is likely your best bet. Agentic workflows allow LLMs to be used for what they are great at (inferring user intent, making decisions, etc.) while outsourcing the things they struggle with (complex calculations) to other parts of a broader system. Text-to-SQL, “tool-use”, few-shot prompting, and other strategies can be combined with real-time data integrations to dramatically improve performance on analytical operations. At Concourse we’ve spent the better half of 2024 getting our AI Agents to be really good at math, and I’m so excited to continue growing our capabilities to tackle even more complex mathematical problems for our customers in 2025. Reach out if you’re interested in learning more about how we’ve built Concourse and the impact we’re already delivering for finance teams.

42

5 Comments

Please media types writing about the cost of AI: recognize there is a difference between huge GPU/TPU clusters needed to train foundation models vs. inference costs to run most apps on top of foundation models. The two are not synonymous, so you can dial back some of the doom and gloom about AI startups and your commingling of the two different AI business types.

11

Love this paper that tests correlation between humans and evaluation models (LLM as a judge). Few big takeaways 1. Cohen's Kappa score is a better measure of accuracy than pure percentage agreement because it takes into account the bias of random agreement/correlation. 2. Humans only agree 97% of the time! This shows how alignment isn't an all or nothing game but rather a distribution where the goal is to get as close as possible without being perfect cause then you risk overfitting against the human labeler who produce the dataset. 3. Open source models are the way to go for evaluation models - performance is comparable (ex Llama3-70B vs GPT-4) and there are so many more benefits from a transparency, explainability, and reliability perspective. Really great stuff from the team at UMass. If you guys have any other cool papers regarding the LLM-as-a-judge space, would love to hear more! Paper Link: https://lnkd.in/edT8P8Tr

85

🚨 New paper w/ interesting insights for fine-tuning and GPU poor folks 🚨 🚀 Exploring Fine-Tuning for LLMs: Full Fine-Tuning vs. #LoRA - Are They Truly Equivalent? 🤔 As LLMs grow, fine-tuning them for specific tasks becomes challenging due to high computational costs. Recently, Low-Rank Adaptation (LoRA) has become a popular alternative, promising similar performance with fewer parameters. But are LoRA-tuned models the same as fully fine-tuned ones? MIT’s latest study dives deep into this question, revealing some surprising differences! 🔍 📝Key Takeaways: 1. Spectral Differences 🧬: LoRA introduces “intruder dimensions” not found in fully fine-tuned models. These alter the model’s weight structure, suggesting that LoRA and full fine-tuning leverage different parts of the parameter space, even if they hit similar accuracy on specific tasks. 2. Generalization & Robustness 🌐: Models fine-tuned with LoRA may struggle with new or sequential tasks, forgetting previously learned knowledge more easily. Full fine-tuning, however, shows better retention and robustness in shifting environments. 3. Effective Rank & Optimal Performance ⚙️: LoRA’s rank matters! Mid-range ranks (like r=64) perform well on tasks and retain pre-trained knowledge, while lower or higher ranks can lead to greater forgetting. Smart parameter tuning (like setting α=2r) can help LoRA resemble full fine-tuning more closely. 4. Why It Matters 💡: LoRA offers incredible parameter efficiency, but its unique trade-offs might affect its use in scenarios needing high adaptability and long-term memory. With adjustments, LoRA can bridge some of these gaps, but full fine-tuning remains the go-to for complex or multi-tasking demands.

25

1 Comment

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

In March, I left Strac (YC W22) to build something new (more on this soon). Here are my key takeaways from taking Strac from zero to one: While at Strac, my favorite part of the role was always taking an engineering-first approach to our go-to-market motion. I loved tinkering, and frankly there's never been a more exciting time to do so as AI rapidly changes how effective sales teams function. Some of my favorite projects at Strac included writing a script to set up emails optimized for outbound deliverability, building a web scraper to check if a company was SOC2 compliant, and integrating our sales tech stack for seamless retargeting and outreach. The most impactful lessons: 1. Sales is a Numbers Game: Five years ago, sales teams relied heavily on traditional methods like cold calling and emailing purchased conference lists. However, nowadays, mass calling and emailing without the prospect showing any intent is ineffective. Today, with the explosion of data available online, intent signals—such as web activity, social media interactions, and content engagement—are critical for identifying leads. The challenge now is not just gathering this data but integrating and analyzing it to make your outbound effective at scale with the right messaging. 2. Get Hacky with Finding Leads & Continuously Experiment: You should run dozens of experiments on different lead lists and iterate until you find message-market fit. During the first month at Strac, we ran over 50 distribution experiments to find where our product had the most obvious signs of product-market fit and iterated on our outreach approach until we found message-market fit. When building out your ETL for sales, it’s important to build a framework that’s conducive to this. If you build a slow-moving, rigid outreach framework, you will limit how many experiments you can run, and your business will die. With startups, every single day you are just slowly dying—iterate fast. 3. Show Immediate Value to Your Customers: Strac was one of the most customer-obsessed companies I’ve ever seen. The most important thing is to fundamentally understand what problem your customer is trying to solve, as there is likely someone else also trying to solve that problem if it's a real problem. Translating this to structuring outreach: if you just make an ask, it's unlikely you'll get a response—you’re not solving any problem your customer faces and instead creating a problem for them. However, if you show someone how easily they can gain something from talking with you, you'll have a prospect excited to have a conversation. As I transition back to the founder role, I'm excited to build tools that address these challenges and help companies accelerate their go-to-market. I will continue to work closely with Strac as they go from one to ten, and I'm excited to have them use our solutions to get there faster 🚀

132

72 Comments

LLMs have become more capable with better training and data. But they haven’t figured out how to “think” through problems at test-time. The latest research from Scale finds that simply scaling inference compute–meaning, giving models more time or attempts to solve a problem–is not effective because the attempts are not diverse enough from each other. 👉 Enter PlanSearch, a novel method for code generation that searches over high-level "plans" in natural language to encourage response diversity. PlanSearch enables the model to “think” through various strategies before generating code, making it more likely to solve the problem correctly. The Scale team tested PlanSearch on major coding benchmarks (HumanEval+, MBPP+, and LiveCodeBench) and found it consistently outperforms baselines, particularly in extended search scenarios. Overall performance improves by over 16% on LiveCodeBench from 60.6% to 77%. Here’s how it works: ✅ PlanSearch first generates high-level strategies, or "plans," in natural language before proceeding to code generation. ✅ These plans are then further broken down into structured observations and solution sketches, allowing for a wider exploration of possible solutions. This increases diversity, reducing the chance of the model recycling similar ideas. ✅ These plans are then combined before settling on the final idea and implementing the solution in code. Enabling LLMs to reason more deeply at inference time via search is one of the most exciting directions in AI right now. When PlanSearch is paired with filtering techniques—such as submitting only solutions that pass initial tests—we can get better results overall and achieve the top score of 77% with only 10 submission attempts. Big thanks to all collaborators on this paper including: Evan Wang, Hugh Zhang, Federico Cassano, Catherine Wu, Yunfeng Bai, William Song, Vaskar Nath, Ziwen H., Sean Hendryx, Summer Yue 👉 Read the full paper here: arxiv.org/abs/2409.03733

180

2 Comments

Batch processing for LLM inference activity: "One sign of that is the rise of batch processing. Today, popular consumer AI products like ChatGPT and Perplexity provide users with near instantaneous responses, otherwise known as real-time inference. However, businesses don't always need immediate responses and are often willing to wait hours, days, or even weeks for responses, as long as they don't have to shell out as much money. Founders of cloud and inference providers have told me there's growing pressure from business customers for this kind of flexibility, otherwise known as batch processing." -- Nathaniel Whittemore on "The AI Daily Brief" (with much deserved shout-out to his inspo from Stephanie Palazzolo at The Information) I fall into this bucket as we have a lot of backfilling to do as we're building out the underlying data in our service, and some of that could just as well be done in the wee hours of the night. The new stuff flowing in I want processed ASAFP, but these backfills—I'd mos def take more cost-efficient delayed batch processing. There's also use cases like needing (near) real-time inference for the first inference pass for timeliness / contemporaneousness value, but then a second pass slow-time inference for wider classification exercises over an ever-growing data corpus. I could see us mixing Sonnet & Haiku models in both cases. Just a heads up to Anthropic Mike Krieger as you contemplate this, along with Amazon Web Services (AWS) Jeff Barr Mark Roy Swami Sivasubramanian on the Bedrock side of things...and I don't mean the "batch" mode that doesn't seem to have any real value other than helping AWS manage its own rate limit challenges. (DeepCast is using Bedrock.)

11

2 Comments