After 8+ years on the Tesla Autopilot team and 3 years at Intel, I started @apexcompute to design a new architecture for efficient AI inference. For the past 9 months, we’ve been building our custom inference accelerator. Today we’re releasing Unified Engine v1. Last June we raised our seed round with @maxitechinc , DeepFin Research, @Soma_Capital and an incredible group of angel investors. In less than 9 months, we completed our RTL architecture and brought our first pre-silicon prototype to life on FPGA. Our architecture combines systolic array and vector processing in a single compute engine with multiple architectural optimizations, achieving very high FLOPs utilization. A single engine is super lean and it uses less than 90K LUTs and 1 MB Block RAM. It may also be one of the smallest logic-footprint compute engines developed so far. Our Unified Engine v1 supports: -matrix-matrix multiplication (~95% FLOPs utilization) -softmax (~90% FLOPs utilization) -broadcast and element-wise operations -RMSNorm / LayerNorm -block quantization/dequantization (fp4, int4) -multi-engine synchronization and many other operations. We even implemented memory-efficient attention similar to FlashAttention, reaching ~90% FLOP utilization. Full benchmarks and the software stack are available on our GitHub: https://t.co/KqTKbB2Inl We have basic compiler written in Python and it supports PyTorch tensors directly to easily test and transfer tensors between the accelerator and host using bf16, fp4 and int4 formats. Our FPGA prototype can already run LLM inference and outperform NVIDIA Jetson Orin Nano, even on a mid-tier FPGA setup (6.4x lower memory bandwidth, 18% slower clock speed at 4.5 Watts). Check the side-by-side comparison video below. Our GitHub includes low-level operator implementations, examples for tiled matrix multiplication, operation chaining, tensor parallelism, attention kernel and a full Gemma 3 1B model implementation. Many more models(Vision Transformers and VLA) are coming soon. Our accelerator IP is AXI-ready for deployment on any AMD(Xilinx) FPGA platform today. Even better, our two-engine prototype runs on an entry-level AMD(Xilinx) FPGA as a PCIe accelerator card. You can purchase it here https://t.co/8B9NOcueVu for $50 to experiment our pre-silicon prototype on your desktop PC or Raspberry Pi 5. We will be releasing hardware bitstream updates as the architecture gets new features. More to come soon! We are expanding our team and looking for compiler engineers and floating-point hardware design engineers. If you're interested, please send me a DM.

This AutoHarness paper (from Google DeepMind) is the most interesting thing I've read lately. I am testing a similar idea (without training) on models like MiniMax-2.5 and getting good results. It already allowed me to synthesize an entire functional coding agent. More soon. https://t.co/pGRRlIEsUR

get PRO on @huggingface and instantly 10x your storage to 1 TB private + 10 TB public ...for $9 a month 😮 a deal this good should be illegal https://t.co/nTMcyTn280

Introducing our biggest upgrade to @googlemaps since the original launch, featuring Ask Gemini (with personalization), Immersive Navigation, and much more!! 🗺️ https://t.co/yjKV44hK6w

// Think Harder or Know More // Chain-of-thought prompting enables reasoning in LLMs but requires explicit verbalization of intermediate steps. Looped transformers offer an alternative by iteratively refining representations within hidden states, but they sacrifice storage capacity in the process. This paper investigates combining both: adaptive per-layer looping with gated memory banks. Each transformer block learns when to iterate its hidden state and when to access stored knowledge. The key finding: Looping primarily benefits mathematical reasoning, while memory banks recover performance on commonsense tasks. Combining both yields a model that outperforms an iso-FLOP baseline with three times the number of layers on math benchmarks. Analysis of model internals reveals layer specialization. Early layers learn to loop minimally and access memory sparingly, while later layers do both more heavily. The model learns to choose between thinking harder and knowing more, and where to do each. Paper: https://t.co/0Gl77zMwOY Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c

Friday the 13th and 10,000 Starlink sats in orbit🔥 @SpaceX is targeting, weather permitting, double-header Falcon 9 launches from the East and West Coast to deploy 54 @Starlink satellites. ... one of these satellites will represent the first time SpaceX surpass 10,000 Starlink satellites in orbit! 🔥

Solar accounted for 54% of all new electricity-generating capacity added to the US grid in 2025. Combined, solar and storage made up 79% of new capacity last year. China has significant growths in solar capacity and generation — the US must be doing the same. https://t.co/uOfdHDTvFJ

The trans school shooter at Tumbler Ridge laid out an elaborate plan to commit mass murder to ChatGPT. A dozen employees saw the plans and wanted to alert police but were turned down by company leadership. Why? https://t.co/9iJ0YscO2H

So basically, we pay the government taxes to fix our problems, and they pay migrants with our taxes to come and cause more problems https://t.co/iRQA3oPneH

What smart idea Sansa, questioning the authority of the Jon in front of all the lords and ladies of the North. Damn Brilliant https://t.co/W6JbgBptwU

Grok Imagine just got a major update from @xai! You can now specify up to 7 reference images for generating a video. The images can reference people, objects, or even keyframes. It's really very flexible! @Grok will then combine the references with your prompt into a video. 📹 https://t.co/9bIDPxOo7m

Grok Imagine UPDATE: You can now integrate up to 5 reference images into your videos. Improved understanding of the prompt Improved consistency between characters and settings https://t.co/OpDIEOlrvq

Grok Imagine now supports Referencing Images with @ for Image Editing and Video on Web https://t.co/R8Uqr2ud2k

BBC says “deep wound” stab kits are now recommended on London buses for some strange, unstated reason https://t.co/9y3cgbLSj4

Historical PSA: We're still in it, more broadly and with privacy, and we will get it right on both fronts https://t.co/Hn5WDPnp6K

i open-sourced autokernel -- autoresearch for GPU kernels you give it any pytorch model. it profiles the model, finds the bottleneck kernels, writes triton replacements, and runs experiments overnight. edit one file, benchmark, keep or revert, repeat forever. same loop as @karpathy autoresearch, applied to kernel optimization 95 experiments. 18 TFLOPS → 187 TFLOPS. 1.31x vs cuBLAS. all autonomous 9 kernel types (matmul, flash attention, fused mlp, layernorm, rmsnorm, softmax, rope, cross entropy, reduce). amdahl's law decides what to optimize next. 5-stage correctness checks before any speedup counts the agent reads program.md (the "research org code"), edits https://t.co/ubU2WPOJwh, runs https://t.co/gJklhqV5hS, and either keeps or reverts. ~40 experiments/hour. ~320 overnight ships with self-contained GPT-2, LLaMA, and BERT definitions so you don't need the transformers library to get started https://t.co/45z8Z7nP3N

Announcing Context7 CLI! MCP isn't the only way anymore. Now any AI agent can pull docs with Context7 — just the CLI and the find-docs skill. One command: npx ctx7 setup

AI isn't replacing programmers, but it *is* making it harder to survive as a programmer with purely technical skills and no interest or expertise in how those skills translate to business or societal value. Funny thing is, this has always been true—it's just being accelerated a bit due to AI. There's a famous essay by @patio11 from 15 years ago called "Don't Call Yourself A Programmer, And Other Career Advice". https://t.co/FlJGJFTWmL

AI doesn’t see the world the way humans do. What looks obvious to us can be misinterpreted by machines when categorizing objects, scenes or context. As AI systems move deeper into areas like safety, healthcare and autonomous systems, these perception gaps become more than technical issues. If AI is going to make decisions about the real world, how do we ensure it actually understands what it’s looking at? https://t.co/PHUybYbKIz @ConversationUS

https://t.co/9WVNz5PdFY great post from @sh_reya, resonates with my experience of ai-augmented writing - it's very much about maintaining momentum and finding your sense of rhythm might just attach this post to my future claude writing chats tbh

https://t.co/9WVNz5PdFY great post from @sh_reya, resonates with my experience of ai-augmented writing - it's very much about maintaining momentum and finding your sense of rhythm might just attach this post to my future claude writing chats tbh

@nospotfer Qwen3.5 is ranked at an impressive 7th place on pinchbench atm. It's really impressive for such a relatively small open-weight model. I think that's a perfect use case for local LLMs. Curious to see where Nemotron 3 Super will rank. https://t.co/Hln4jMPSi8

🚨BREAKING: Langflow just open-sourced a complete RAG platform in a single package. It's called OpenRAG. Built on Langflow, Docling, and OpenSearch. Upload documents, run semantic search, and chat with your data no duct tape required. One command to run: uvx openrag → Full document ingestion pipeline → AI-powered chat over your files → Built-in semantic search via OpenSearch → Visual workflows via Langflow → Docker support out of the box 100% Opensource.

Context engineering is having a resurgence as a topic and I'm here for it @LoganMarkewich and I wrote about this a while back. The short version of what we think actually matters 👇 The framing: prompting is just one input. context is everything the LLM sees: system prompt, chat history, retrieved knowledge, memory. Context engineering is about engineering all that, well.. → context compression/summarization before it hits the LLM → relevance ordering, not just relevance filtering → being deliberate about what goes in long-term memory vs short-term Our full post: https://t.co/zdNDk87k8j

Another week, another noteworthy open-weight LLM release. Nvidia’s Nemotron 3 Super 120B-A12B looks pretty good. Benchmarks are on par with Qwen3.5 122B and GPT-OSS 120B, but the throughput is great! Below is a short, visual architecture rundown. https://t.co/eqSOPAeqQz

Bitsandbytes now has full ROCm/CUDA feature parity! 🧵 See thread for more details. https://t.co/ZeqRzkuXIb

Bitsandbytes now has full ROCm/CUDA feature parity! 🧵 See thread for more details. https://t.co/ZeqRzkuXIb

MCP sucking is a harness problem, not an MCP problem MCP unlocks behavior that is fundamentally impossible to get via CLI or APIs Bad auth, too much context usage, all get solved with an execution layer - your agent writes code to progressively discover and call tools https://t.co/kUfhhEkiph

Learn more and apply from the link below🤗 https://t.co/QLPgege4CI

anyone else notice that chatgpt is using clickbaity follow up questions? https://t.co/S7wTlYf38e

Hugging Face Builders is a global community program that puts local leaders at the center of the open-source AI movement 🤗 If you're passionate about open AI and love bringing people together, this is your invitation to lead ✉️ Apply for to build the Paris chapter today ➡️ https://t.co/ONVBZdxRdc