Ceramic.ai

Our CEO and founder Anna Pattersonis joining Cerebral Valley AI Summit to lead a discussion exploring the next wave of progress in AI performance: improving efficiency at the infrastructure layer. As infrastructure quickly defines its competitive edge, Anna will discuss how we can push the boundaries of what AI can achieve. 📅 Don’t miss her session on November 12th at 12pm PT: https://lnkd.in/ggdz3yKP 

Proud to see our founder/CEO Anna Patterson has been named to the 2025 Mayfield | Divot AI List, recognizing 50 builders shaping what’s next in AI before the hype and headlines. We’re honored to see Ceramic’s work highlighted alongside so many brilliant founders, researchers and operators. Big thanks to Mayfield and Startup Grind for including Anna in this year’s list. 🔗 Check out the full list here: https://www.divot.org/list

If you wanted to kill ChatGPT, how would you do it? Turn off thousands of servers or corrupt terabytes of memory? It turns out, it may only require changing a single 1 to a 0… This is possible due to "super weights" found by researchers at Apple—single neurons so critical they act as a kill switch. The effect is dramatic. For example, a healthy Llama-7B responds to "My favorite condiment is..." with "mustard." But with one super weight zeroed out, it says: ":/νώ好 !\β.” That was an older, dense model. We asked: are today's advanced Mixture-of-Expert (MoE) models immune? Ceramic.ai's latest research, done by intern Andrew Wesel, reveals a split. MoEs from OpenAI and Allen AI are robust, but we found Alibaba's newest Qwen3 models still contain these critical points of failure. This has major implications for model security, compression, and our basic understanding of how LLMs work. See the data for yourself: https://lnkd.in/gWhvHF-9 What model should we test next?

OpenAI buried 5 cryptic constants in their SwiGLU code, but when Ceramic.ai let each transformer layer choose its own activation recipe, we uncovered a shocking truth 😲: forcing all layers to use identical activations is like making everyone wear the same shoe 👟 size - wildly inefficient and easily fixed 🛠️ for a 20% performance boost 🚀. https://lnkd.in/gZkHXHuJ

Our Founder/CEO got to sit down with Riviera Partners and talk about one of her favorite topics ... exploring the state of AI! #foundertalk #venturevanguard #stateofai

Ceramic.ai we're writing this negative result up for posterity and decided to share it. We discovered a Ferrari that only works in school zones … 🚗💨 The math checks out beautifully - decomposed parameters learn faster! Except they only work at learning rates 📈 so conservative your model trains slower than dial-up internet 🛜 . Push it to production speed? Instant death. Classic academic trap: optimizing for elegance instead of shipping. Sometimes the "dumb" solution wins because it doesn't explode when you need results yesterday. #AI #Shipping #RealityCheck Read the full blog here: https://lnkd.in/exNW5gnM

Here’s the brutal reality: If AI labs were chip fabs, then we are officially in the early days of manufacturing - working on how to make a product more efficient without revenue loss. Here Ceramic.ai our Chief Scientist, Tom Costello, looked at yield rates and causes to see why AI labs are burning through GPUs at high rates. take a look --> https://lnkd.in/g3UPj9Ak

Training large language models efficiently at scale is hard — Ceramic’s architecture makes it easier. By combining breakthroughs in mathematical optimization, network efficiency, and long-context specialization, Ceramic delivers high-performance training with fewer resources. 🚀 In our latest test using Lambda’s new B200s, Ceramic trained long-context LLMs on just 8 GPUs — and outperformed every benchmark we’ve seen. 📊 See the full POC results: https://shorturl.at/3HkDZ 🔍 Learn more about our technology: https://lnkd.in/g3Dx3G6b

Surprise! Hugging Face's default settings have been quietly sabotaging your LLM evaluations - your eval loss drops as batch size increases, which is odd since you're testing the same data. The "fix" is simple and will actually speed up your evaluation by 50%+ ... take a look at what our intern YU-CHI HSU discovered in our latest blog! https://lnkd.in/gKG2jY7J

Our chief scientist, Tom Costello, decodes the rise of Chinese models. https://lnkd.in/gaE3ZmXj Happy 4th! Happy reading.