Christophe Ponsart

ContextFit just hit 99.0% retrieval on LongMemEval-S n=500. The unlock: memory atoms + fusion. Instead of treating memory as one flat vector search problem, we route preferences, decisions, temporal updates, and open loops differently. Agent memory needs structure, not just more context.

Got my first GitHub star today. Tiny number. Weirdly meaningful. I’ve spent my whole career around consulting companies, where most of the best work happens behind client walls and never becomes open source. So seeing someone outside the room find a project useful enough to star it feels different. Small milestone, big feeling.

I’m starting to write publicly about ContextFit, an open-source project I’ve been building around token-native memory retrieval for AI agents. First piece: why I think AI memory needs to move beyond “embed everything and hope the nearest chunks are enough.” AI memory should be fast, local, auditable, and built for the way models actually consume information: tokens.

There is likely not much news this year that is bigger than this (anticipation is real). Jony Ive’s designs have changed the way the world works. iPhone and MacBook are some of the most successful products in history and created the largest market cap company in the world (Apple). They were also significantly different from anything the world had seen at the time. I suspect and hope the product reveal next year will be just as revolutionary and magical. There’s nothing I find more exciting than to be surprised and get to experience a new product that changes the way you work and brings moments of pure joy and imagination. I am looking forward to the reveal! xitter.twitterapi.xyz/altryne/status/1925235…

Help me find my next book. Any recommendations? I’ve never been much of a reader (maybe it’s thanks to Audible… is that still reading?) but I can’t seem to stop consuming information the last couple years. Something about this AI race has me consuming like I’ve never consumed before. Here are my recent reads, which I highly recommend: - Co-Intelligence by Ethan Mollick - The Coming Wave by Mustafa Suleyman and Michael Bhaskar - Masters of Scale by Reid Hoffman - Chip War by Chris Miller Anything in the same genre would be great. Cheers!

xAI Grok 3 announcement tonight will be one to watch! 8pm PT DeepSeek threatened the thesis that large GPU clusters are required to create advanced AI models. This led to the largest company one day loss of market cap in history. xAI has the largest computer cluster in the world with 100k GPUs. They have the talent, they have the data and they have the compute. This is quite a recipe and something to keep an eye on… There are many ways to create advanced models and there is room for significant innovations and improvements as we are in early stages. But, I think it will continue to become clear that more compute will continue to lead to more capable models! My 🍿 is out! Who’s watching with me?

DeepSeek - the ultimate Trojan Horse AI feels like it’s turned the corner of being considered “hype”, and become a race for power and supremacy. Competition is in our nature and consumers of AI have accepted the risks of trusting companies like OpenAI and LLM technology in favor of the benefits and efficiencies it provides. The challenge with this race is that you become laser focused on winning and you stop evaluating risks as closely. A couple weeks ago, TikTok was banned in the US (and unbanned within 24hrs) due to data security risks of the app collecting sensitive data and its influence on the nation. A week later DeepSeek R1 was released as open source and made available in Apple and Android app stores. Within a few days major technology cloud vendors including Microsoft Azure deployed the model on their infrastructure. The DeepSeek apps have significant data privacy and security risks. Since Large Language models respond best with more context (I.e. information), users are drawn to provide more details about themselves, their companies and their specific circumstance. This could include, geographic location, health data, company sensitive process and/or intellectual property. Clearly a stark contrast from the data collected from teenagers vigorously scrolling through videos on TikTok. Many have warned about these risks, yet the iPhone app made it to #1 ahead of chatGPT on the marketplace. Tech companies and engineers are well aware of this issue and are opting to deploy the open source version of the model on their own servers. This protects user data as all data and processing never leaves your environment. But, does this remove all risk? The challenge with Large Language models is that they act like a black box. Even if you host your own LLM, it’s virtually impossible to understand what information the model has been trained with and how it will respond to various requests. Hence, the ultimate TROJAN HORSE! We should not just be concerned about our data leaving our servers. Maybe, we should also be mindful of the data being delivered at our doorstep within LLM black boxes! Early OpenAI models censored responses, Google Gemini steered cultural biases and Elon xAI set their goal to understand the true nature of the universe. What are DeepSeek’s core training principles? What is being censored and what is being cleverly interweaved into the fabric of its knowledge and reasoning model? Engineers are using LLM models to generate software code; Writers to write their stories; Designers to extend their creativity. LLM models are going to be part of almost all content we generate as a society. The weights in our LLMs will steer the content we create, influence our society’s train of thought and be embedded in the automations of our systems. There are many paths here in which hidden viruses can be unwillingly introduced. It is imperative we take careful consideration in which LLMs we use as well as the motivations of those that are developing them. LLMs act like a black box. If the outputs are challenging to validate you need to build trust with their input!

Headed to CES from gate “infinity” ♾️ this morning - seems appropriate. 🤣 Felt inspired to go this year as AI will give massive opportunities to consumer electronics from both a software and hardware perspective. AI’s ability to bridge unstructured real world information will lead to new ways to engage with computers that feel more natural and seamless. Looking forward to seeing what the world of electronics has been working on this past year and what it envisions as our future!

Is AI in a bubble? I throw some punches and blow myself up with genAI! It's time to find out. Also, Nvidia announcements tonight... 👀

10/10 RoboTaxi event.. What if Cyber cab has a front trunk designed for food delivery? Keeping items warm, cold and why not drink holders! Should we call it a FrInsulator?? Lowering cost of transportation is going to greatly increase demand! We are going to see a significant ramp up of vehicles on the road delivering significantly cheaper items like food. This is not a linear growth 👀 Also, the desire for “item” delivery will outpace the traditional taxi (people mover). For all these reasons, you need to create a better food/item delivery experience. This affects the vehicle (insulating food) and also requires partners like McDonald’s which are global to be in loop through technology (Tesla fulfillment app) and process to know how to service food to a cyber cab (open/close/lock front trunk). Can’t wait for 10/10! @wholemars , what do you think?

Cybertruck FSD incoming… let’s go!

HOW did NotebookLM make consulting service offering slides actually interesting to listen to?? I uploaded our "Applied AI practice" presentation deck and it created this entire podcast automatically 👀 I just can't stop listening... HOW? I do have to say that it is way too "Fascinated" by typically less important facts ("Their project app process is fascinating 🤣 ") but it does such a great job of coming up with very simple analogies for complex topics! Wow!

@SawyerMerritt @wholemars @farzyness Elon referred to Robotaxi as Cybertaxi (before correcting himself) on quarterly call, yet no one is talking about this… I’ve assume robotaxi would have a cyber skeleton for durability and other benefits but not sure it’s ever been officially announced, has it?

🚀 Highlight of last week: A firsthand experience at the @Microsoft AI Tour in NY alongside @vanessafiola . The realm of possibilities with AI is vast and, as emphasized by @satyanadella , we’re encouraged to “Build the most ambitious thing you can think of – and cost optimize later.” In a world where #GenAI is reshaping media and CPG industries, here are some of my insights: Microsoft is ALL IN on GenAI: having access to the leading transformer models from OpenAI, they are scaling the entire technology stack to leverage and scale LLMs. From chips, to cloud, to Windows OS all the way to integrating them within their apps Compute is expensive and Chips are constrained: LLMs are compute intensive and chips that can run them effectively are in high demand and limited quantity. Microsoft efforts are on all fronts including partnering with vendors outside of NVDIA, designing their own chips and most importantly making more efficient use of the resources in place today. Lots of work being done on optimizing workloads for max capacity usage of existing chips Change Management unlocks the Enterprise: every company is looking to reduce costs and create efficiencies by automating workflows and empowering their teams with unfettered and instant access to their enterprise data. GenAI is an enabler to realize this vision but will require end users to work different and developers to think different in how we engage with our users. Change has and will always be difficult for users and properly managing this is key to unlocking enterprise effectiveness. We’re actively hearing this at @qvest_us from Fortune 1000 brands. It’s time to educate, collaborate and deliver AI is fundamentally altering our lives and work. What are your thoughts on the intersection of AI innovation, cost, and change management? How do you see Gen AI influencing your industry? #AI #MicrosoftEvent #Innovation #DigitalTransformation #AIforGood #QvestUS What would you add?

They say, when life give you lemons, make lemonade. I say when your family gives you a bell, you make a cybertruck 🤣 How did I do @woodhaus2 @elonmusk ?