Typesafe.ai, founded by former OpenAI researcher Diogo Almeida (@CompleteSkeptic) who helped invent Reinforcement Learning from Human Feedback (RLHF), just released a model that is ~200x faster and 450x cheaper than leading LLMs, while matching near-frontier performance on decision-making (System 1) tasks. It hits that speed and cost profile by processing data in parallel rather than generating token by token which is how most current models like Claude or ChatGPT function. Most importantly it outputs confidence scores when it’s uncertain meaning the model tells the user how confident it feels in the decisions it is making. This will become increasingly more important as AI regulations pick up over concerns about safety. Why this matters: most knowledge work is about decision making not text generation. Software that people use has always just been static workflows. Every time you click a button, you've made a decision. When you click “Send Email” the software just executes the fixed path of actions you committed to that some developer has programmed in so you can send an email halfway across the world. When people talk about AI replacing the “human in the loop” what they actually need is a system that can evaluate choices and quantify risk in real time: Should we flag this transaction? Should the car turn right? Does this patient require surgery? Fast, dirt cheap decisions paired with calibrated confidence scores are the missing link for real-time agentic workflows - especially right now, when ROI is the single biggest question hanging over the AI buildout. We have seen open source models pick up a lot of steam as companies have tried to lower costs and delegate simple work to open source and high level work to the frontier. I think Typesafe AI’s model integrates nicely into this future providing cheap and fast decision making as well as providing a way to monitor the confidence of the decisions being made for when we still need that human touch. In order to achieve AGI, the goal is to first create a computer that thinks like a human. The human brain runs on about 20W of power and it has gotten us from sticks and stones until now. AI uses ~1000W. That leaves a lot of room to be made on improving AI models. The human brain provides decisions but also a confidence in those decisions. You may believe the Eagles are going to beat the Chiefs in the Super Bowl, but the level of confidence you have in that answer matters. Depending on your confidence in the Eagles ability to win, you may feel more or less nervous betting those $5 with your friends. Furthermore, no human thinks one word at a time. We think in concepts and then figure out how to express ourselves. We also have saved behaviors (call it muscle memory or reflexes) that trigger when someone throws us a ball, tells us to turn right etc. rather than involving a multi-step reasoning process every time. I think JEV (Typesafe.ai’s model) and the work on looped transformers represent a big step in the right direction towards AGI.

The market is undervaluing @SakanaAI’s DiffusionBlocks (ICLR 2026). As CapEx spend becomes an increasingly larger worry for investors, diffusion blocks allow model makers to avoid sinking as much CapEx into top-tier chips just for VRAM capacity and split some of the training load across older racks. Rather than needing 1 Big GPU with lots of ultra-high-bandwidth memory you can instead split the load across 4 medium sized GPUs with smaller memory footprints. In the 2000’s, rather than having to buy 4 1GB modules for your database to keep things running you could instead buy 4 256MB modules for 4 separate servers. While distributed systems ultimately led to more total demand it did put pressure on the premium memory as the market expanded and things commoditized. While inference memory has seen aggressive optimization (@MIT @Tsinghua_Uni @deepseek_ai @Kimi_Moonshot), training remains constrained by the realities of training a LLM. Training is expensive because deep backprop creates a brutal HBM bottleneck. Holding parameters, gradients, optimizer states, and activations across an entire model forces you to buy huge GPU clusters just for the memory. DiffusionBlocks splits an N-layer Transformer into B sub-blocks of size N/B trained independently on separate GPUs with significantly reduced inter-block gradient communication. To be clear, memory and compute requirements remain the same. But in the same way distributed systems in the 2000’s helped reduce the premium on the ultra-high memory at the time, I believe DiffusionBlocks could pose the same problem. Active state memory drops by 1/B per node (66% - 75% reduction): Peak VRAM requirements per node collapse while total parameter, memory, and compute capacity stays constant. Pipeline Parallelism drops significantly: Eliminates cross-chassis InfiniBand/RoCE bottlenecks and pipeline bubbles. Tensor/expert routing stays local over NVLink, pushing MFU from 45% → 55% (~18% real cost cut) with standard autoregressive inference preserved. Legacy hardware unlocked: Older GPU racks move beyond basic non-agentic tasks and can actually participate in pre-training. Downside: $NVDA, $SKHY, $MU, $SNDK and Samsung profit from network depth requiring monolithic VRAM. Decoupling the two drops a 2.2T MoE’s memory training floor from 128 Rubin GPUs to 48 (a single NVL72 chassis), pressuring premium memory margins. You can use less memory per chip as you can independently train blocks of the model. Upside: Total compute (C ≈ 6 • ActiveParameters • Depth) and memory is conserved. In reality this just distributes the same work just across more computers. By lowering the barrier to entry, training becomes cheaper and more efficient and it becomes more economical for companies to train their own models for their own use cases. Furthermore, inference remains a persistent problem that still needs that premium memory. As cost per token drops, Jevons Paradox kicks in and cheaper training unlocks new positive-ROI use cases and explodes aggregate token volume. While this technology is still in its infancy, I think it is important to note that like most innovations, it takes about 6 - 9 months for the technology to be implemented, adapted, and rolled out at scale.

I did research on @deepseek_ai recent papers and the implications for the R2 model later this month. I think Deepseek has built a ~10x more efficient model and they will offer a better AI than OpenAI at a similar price or a similar one at a cheaper price. warrenniles.com/research/202…