Fsbonetto (@reasoningTokens) header
Fsbonetto (@reasoningTokens) profile picture

Teaching LLMs to think ahead; Learning about ML since 2017

Joined September 2017
Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)Photo by Fsbonetto (@reasoningTokens)
Replying to @karpathy
@karpathy did a great job with auto researcher. Using a similar harness AI is able to build a @risc_v CPU better than humans using @YosysHQ: github.com/FeSens/auto-arch-…
3
437
Fsbonetto retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
38,249
79,720
17,442
716,135
24,404,798
Can we make this get to @elonmusk A proposal to improve @SpaceX's Starship fabrication.
2
252
Replying to @claudeai
@claudeai Fable 5.2 routed HDMI 4k, DDR3, USB hi-speed, FPGA. Optimized BOM cost reducing 44 dollars of each board made. Validated the board with electrical simulation, found erros in the footprints, voltage drops, fixed all of them. 6 layer board. @AnthropicAI cooked with this one. Going to send to @JLCPCB and tell you the results soon.
6
3
1
18
1,564
Half of it was made with @BackplaneWorks but I felt like it hanged too often and consumed too much tokens on the PCB routing. Worked perfectly on the schematics though.
2
477
Fable 5.2 just routed a PCB with DDR3 memory. Vibe engineering is becoming possible
1
1
4
826
OpenTPU got a real server that you can use with @opencode ! What started as a toy project now is able to serve Qwen 35B at 30 tok/s and do real work!
1
5
309
We ported openTPU's ISA simulator to @bendlang . Identical to our Python reference, 10–15× faster on real models (Qwen3, Qwen3.5, LFM2), ~60× on the MoE expert-cache sim. 110+ laws checked by Bend's proof checker, and the port found 2 bugs in our Python sims! Thanks @VictorTaelin for the awesome language. github.com/FeSens/openTPU
1
2
200
Fsbonetto retweeted
Hand-tuning a CPU pipeline means endless cycles of guess, rewrite, re-verify and re-run place-and-route. auto-arch-tournament by @reasoningTokens automates that loop: an agent proposes microarchitecture changes to a SystemVerilog RV32IM core, then checks them with riscv-formal, Verilator cosim and CoreMark CRC. It runs on Codex CLI or Claude Code, is Apache 2.0, and scores fitness with 3-seed place-and-route on a Tang Nano 20K.
1
1
1
73
CPU demand for agents is real. I my experience 6x more flops are needed than before. Before I had one MacBook and was happy. Now I have 2 MacBooks, 2 Desktop PCs, 1 server with 20 cores and 128GB or Ram and it still not enough. The agent demands more.
1
130
AI agents designed hardware that runs Qwen on an old FPGA github.com/FeSens/openTPU. @aviflombaum sees a bigger story: what tiny, fast AI hardware could mean for medical devices.
1
1
3
85
I just ran an LLM on an FPGA. Now I'm wondering: how far are we from building our own TPU? Not in simulation. Not on a GPU. On a real Xilinx Kintex-7 FPGA with 4GB RAM. A few days ago, I came across openTPU, an open-source AI accelerator project created by Felipe Sens Bonettoe with the help of AI agents. The idea was too interesting to leave on GitHub. So I decided to put it to the test. 🔧 From GitHub to real hardware: → Connected the FPGA board to MINISFORUM PC via PCIe → Configured the hardware → Built the software stack → Brought up the DMA drivers → Loaded real model weights → Started testing inference And it actually worked. ✅ FPGA hardware detected and initialized ✅ PCIe DMA communication working ✅ Memory subsystem operational ✅ Hardware self-tests passing ✅ Real LLM inference running, including Qwen The most interesting part wasn't the performance. It was realizing that you can take an open-source accelerator architecture, bring it onto real hardware, and start experimenting with your own AI inference engine. Of course, this isn't an H100 competitor. And an FPGA prototype is still a long way from a production ASIC. But that's exactly why I find it exciting. 🔬 What's next? A few years ago, building your own AI accelerator sounded like something only a semiconductor company could attempt. Today, open-source hardware, affordable FPGA boards, and AI-assisted engineering are making that idea worth experimenting with. Maybe the next wave of AI innovation won't just be about building better models. Maybe it will also be about who gets to build the hardware that runs them. 👏 Huge credit to Felipe Sens Bonetto for creating openTPU. teapawt.org/reasoningTokens 🔗 Original project: github.com/FeSens/openTPU #FPGA #AIHardware #OpenSource #ASIC #LLM
6
3
15
445
Will we ever see @Etched benchmarked on @SemiAnalysis_ Inference X? It would be very interesting
2
156
Open Inference on Open Hardware. #OpenTPU is now able to do @Alibaba_Qwen 3.5 inference at +30 tok/sec! And a 44% increase in prefill speed driven by fully automated agents research in the last 24h! github.com/FeSens/openTPU
6
159
Turns out that AI is now capable of designing its own hardware: 25 tok/s on Qwen 3.5 github.com/FeSens/openTPU
2
3
1,777
Fsbonetto retweeted
Starting from wallets, gas and transactions - the foundations every Ethereum builder needs. Thank you @reasoningTokens!
Replying to @eth_floripa
🧩 Chain, Wallet & Gas - Felipe Bonetto @reasoningTokens former Head of Data Eng at VAAS, where he built graph tech over 100B+ rows of onchain data to catch fraud and scam wallets, stopping 4,000+ malicious txs. He's also a software engineer, AI researcher, and now a founding engineer at a stealth silicon valley startup.
1
10
318
I’ve being poking around with agent harness and fully autonomous companies and I think I’ve seen end game
4
134
CLI tools is quickly becoming the standard for agents. If your company don’t have a cli tool, you will be left behind
3
128