CTGT (YC F24)

10 years ago, I was a kid watching Google I/O amidst periodically disconnected household utilities. I was totally captivated by the nascent applications of AI on display and trained rudimentary models on the aging hardware (an old laptop) I had access to. Today, I'm excited to announce that CTGT (YC F24) has raised $7.2M led by Gradient, Google's early-stage AI fund, to help enterprises scale AI beyond deep learning. The round is joined by distinguished investors including General Catalyst, Liquid 2 Ventures, and Y Combinator. We’re grateful to be supported by luminaries in AI, including Fran?ois Chollet (Keras), Paul Graham (YC), Peter Wang (Anaconda), Michael Seibel (Twitch), Mike Knoop (Zapier) and Wes McKinney (Pandas). We believe this is the most important problem to be working on today. Here's why: Since my undergraduate work, I’ve been obsessed with elucidating AI's unyielding demand for compute. AI models keep getting bigger, but the fundamental inefficiencies of deep learning remain. DeepSeek showed us that hyper-optimizing model training can push performance further, but scaling alone won’t fix the underlying problem, especially as timelines to AGI are hotly contested. This is what CTGT is solving. We've built a new kind of AI stack – one that removes the constraints of traditional deep learning by rethinking how models learn and train. It customizes, trains, and deploys models up to 500x faster with state-of-the-art accuracy on a wide variety of tasks. All without requiring massive compute. Our AI deployment and quality platform has already been used by Fortune 10 enterprises to gain more control over their AI models in real-world environments. Now, we’re expanding access to more enterprises looking to move AI from proof-of-concept to production. This is just the beginning of our journey in creating the next generation of truly intelligent AI: built from the ground up to be trustworthy and efficient, dynamically adapting to your needs. If you're interested in working at the forefront of intelligence, join us. And thanks to all our supporters, including Kulveer Taggar, Adil Syed, Nate Matherson, Jake Mintz, Matthew Lenhard, Character Capital, Deepwater Asset Management, CoreNest Capital

Thanks Alex Heath for featuring our work at CTGT (YC F24) on removing DeepSeek-R1 censorship in The Verge! DeepSeek-R1's censorship severely limits its potential by blocking valid responses and reducing flexibility. Feature-level intervention solves this problem. Post-training fine-tuning often creates over-censorship, introducing inefficiencies. Perplexity's R1 1776 showed how dataset-based post-training can override restrictions, but these approaches remain fixed and require manual work. A feature-level intervention framework directly modifies internal activations responsible for censorship in DeepSeek-R1-Distill-Llama-70B, offering a more efficient solution while maintaining reasoning capabilities. Experimental results: - Baseline model answered 32% of sensitive queries. With intervention, response rate reached 100%. - Accuracy across reasoning, mathematics, and coding tasks remained statistically unchanged. - Runtime overhead was negligible. This approach works across different LLM architectures. By modifying internal representations tied to censorship, we can get real-time, precise control over model behavior without retraining.

What takes 6 months in San Diego happens in 1 week in San Francisco. During our fundraising round after YC, my co-founder Trevor Tuttle pointed out the San Diego investor who scheduled months out despite having allocation. Every SF investor wanted to meet immediately. We closed our entire round in less than one week in SF. Investors created unauthorized calendar meetings. Some showed up at our door unannounced. This couldn't have happened anywhere else. The velocity of SF's tech ecosystem directly translates to execution speed. The compressed timelines allow founders to build momentum that would take months to establish elsewhere. This environment becomes particularly valuable when building foundational technology. Investors understand urgency and reward conviction with capital that moves at the pace of innovation. The San Francisco advantage is real.

AI models are beginning to hit the limits of compute. Model size is far outpacing Moore’s Law, training runs for large models can cost millions of dollars, and we’re stumbling around in the dark: finding methods that work well, pushing them to exhaustion, and then grappling with issues like hallucinations after AI is widely deployed. DeepSeek’s R1 has proven that throwing more compute is not the only path forward. It optimized training costs but still remains firmly in the deep learning paradigm, constrained by its scaling laws. This is “The Bitter Lesson” in action. Once Nvidia’s CUDA enabled efficient tensor operations on GPUs, deep networks like AlexNet drove unprecedented progress and homogenized the field into a compute-heavy approach. But can we really reach artificial general intelligence (like the AI in Blade Runner or 2001: A Space Odyssey) simply by adding more parameters to these LLMs and more GPUs to the clusters they are trained on? My work at UCSD was predicated on the belief that this scaling will not lead to genuine intelligence. Unlike other sciences that reconcile practice with theory, such as Maxwell’s equations in electromagnetism, we have grown accustomed to brute-forcing AI. Yet it does not have to be this way. By isolating the feature-learning process integral to deep neural networks, we developed a backpropagation-free AI stack that removes the computational excess of deep learning. The result is far greater efficiency and interpretability, without endlessly throwing GPUs at the problem. As Ilya Sutskever points out, everyone is searching for the next big thing. Perhaps we should stop ignoring “The Bitter Lesson” and strive for a theory of everything in learning. That is exactly what we are working on at CTGT: building AI from the ground up with a principles-first approach that rethinks how learning actually happens.

It’s podcast time! ?? Check out the latest episode of the TechCrunch Equity Podcast where Mary Ann Azevedo sits down with Eylul Kayin, partner at Gradient, to chat about the AI startup landscape. Eylul was recently a judge for the 2024 Startup Battlefield and shares her advice for founders applying, as well as a few fun facts about the history of the program. It was also announced last week that Gradient is leading the $7.2M seed round of our Startup Battlefield Top 20 alumni CTGT (YC F24) ? Tune in next week when I will be on Silicon Valley Impact talking about how founders make decisions. Decisions like- Should I apply for Startup Battlefield? ?? That’s an easy one, yes you should! Applications are open now, head to TechCrunch.com/apply to learn more #StartupBattlefield #Disrupt2025 Brian?? Sparkes Trevor Tuttle Cyril Gorlla AJ Thomas ?? Michelle Florendo

I analyzed 100+ pre-seed and seed startup websites in the last 3 month. Here's what I found. Most are making the same mistake. I admit, I have done it too… Early-stage startups try to copy late-stage company playbooks. They use "aspirational messaging" that ends up saying absolutely nothing: "AI-powered solutions for transformative growth." ?? Problem? Nobody understands what you actually do. Even if you solve their problems! The startups that are getting it right focus on three simple things: 1. What value do you give customers? 2. Who is it for? 3. How does it make their life easier? (Show the outcome) That's it. If someone can't explain what you do in one sentence after visiting your site, you're losing customers. Check out how Resend, CTGT (YC F24) and Fabi.ai has done it. ↓ What's the clearest startup website you've seen recently? Tag them below. I'm always looking for great examples.

Now that the dust of DeepSeek's R1 has settled, what are the implications for how we approach AI at large? First principles thinking is essential to building the next generation of AI. Thanks to Fran?ois Chollet, Dalton Caldwell, and Rajesh Gupta for their comments.