RootFTrust @parallel_beings

Joined August 2009

Tweets

159
Followers

215
Following

1K
Likes

3K

RootFTrust @parallel_beings

4 months ago

I now feel the gap to be filled is when LLM have access to hardware around it. It can be an IP camera or a single board computer or router to more serial communication devices. Now as more people get into vibe coding the hardware access becomes more important. For example we gave agents access to serial programming tools to program other IP and non IP devices. respcode.com

0 0 0 53 0

View Details

RootFTrust @parallel_beings

4 months ago

respcode.com/blog/bare-meta…

0 0 0 108 0

View Details

RootFTrust @parallel_beings

4 months ago

Using Respcode Hardware Agent we asked 5 frontier AI models to write bare-metal Ethernet UDP firmware for the @Arm based @ST_World Nucleo-F767ZI without feedback loop; no HAL, no lwIP, no SDK. Using Raw register access to drive an LAN8742A PHY over RMII, build UDP packets from scratch, and handle ARP. In Compete Mode, only 2 of 4 compiled. Only 1 sent real packets on the wire. @AnthropicAI Opus 4.6 and Sonnet 4.6 ran on real HW. Every binary flashed and verified on real silicon. For technical details check the blog and Link in the comment.

1 0 0 232 0

View Details

RootFTrust @parallel_beings

4 months ago

Bare-metal Ethernet is where "hello world" goes to die. No lwIP. No HAL. Just you, the STM32 reference manual, and 1,400 pages of register definitions. To get a single ping, you have to: → Initialize the RMII PHY (LAN8742A) and nail the MDIO timing → Configure MAC DMA descriptors with strict memory alignment → Build Ethernet frames (CRC and all) from raw bytes → Handle ARP requests and TX/RX ring buffer management Most engineers spend days debugging a "silent" wire. We asked 4 frontier models to do it in under 60 seconds. RespCode's Compete Mode ran Claude Opus 4.6, GPT-5.3, Gemini 3.1 Pro, and DeepSeek Coder on the same prompt. The results were compiled in our cloud sandboxes and flashed to a real Nucleo-F767ZI. 📊 The Results: 🏆 Claude Opus 4.6 — 7,964 bytes of lean, working firmware. Real UDP packets on the wire. ARP responses confirmed via UART. First try. ✅ Gemini 3.0 Pro — Compiled and ran, but required a slight buffer tweak to hit the wire. ❌ GPT-5.0 & DeepSeek — Hit linker errors — proving why you should verify, not trust. The process was entirely autonomous. Prompt → Hardware in under a minute. No manual debug loop required. We're building RespCode to make firmware development as fast as web dev. 11+ models competing, real-world hardware verification, and an agent that handles the "flash-and-fail" cycle for you. 🚀 Try the Free Web IDE → respcode.com #EmbeddedSystems #Firmware #STM32 #Ethernet #BareMetal #AI #DeveloperTools #RespCode

0 0 0 232 0

View Details

RootFTrust @parallel_beings

4 months ago

@arvindgunasekar

0 0 0 34 0

View Details

RootFTrust @parallel_beings

4 months ago

@oprydai the urge to read every page of a document is strong because the cost of an error feels high. However, the most effective learning happens in the "struggle" of the implementation.

0 0 0 17 0

View Details

RootFTrust @parallel_beings

4 months ago

We asked 4 flagship AI models to write a simple bare-metal firmware for the same board. 3 compiled. Only 1 actually worked on real hardware. Gemini 2.5 Pro: LED blinking ✅ — Confirmed on real silicon. The green RGB LED (PIO1_4) toggles at ~4Hz, exactly as predicted. Claude Opus 4: Dead ❌ — No LED activity. GPIO writes go to 0x40090xxx (unmapped), as predicted. DeepSeek Reasoner: Dead ❌ — No LED activity. GPIO writes hit byte-register area instead of DIR/SET/NOT, as predicted. Technical blog - respcode.com/blog/how-respc…

0 0 0 103 0

View Details

RootFTrust @parallel_beings

4 months ago

How RespCode Compiles Real Bare-Metal Firmware with multiple code models. Here's what actually happened when we asked Claude Sonnet 4.5, DeepSeek Coder, GPT 40, and Gemini 2.5 Flash to write a bare-metal LED blink for the LPC55S69-EVK board. The SVD validator reveals exactly which models know this chip's peripheral layout and which are guessing. Next step is to validate the firmware compiled on real hardware. respcode.com/blog/how-respc…

0 0 0 116 1

View Details

RootFTrust @parallel_beings

5 months ago

Also asked 4 different models to develop a basic crc32 test program.

0 0 0 67 0

View Details

RootFTrust @parallel_beings

5 months ago

Asked four different models to write a memory bandwidth test for x86 and Arm64, only one compiled without any error. respcode.com

1 0 0 83 0

View Details

RootFTrust @parallel_beings

5 months ago

@EHuanglu That's the reason we built respcode.com the models compete/collaborate with each other to produce the best results.

0 0 0 35 0

View Details

RootFTrust @parallel_beings

5 months ago

@never_released hi

0 0 0 24 0

View Details

RootFTrust @parallel_beings

5 months ago

@JustinLDallas Would you like to test verilog programs using respcode.com?

0 0 0 20 0

View Details

RootFTrust @parallel_beings

5 months ago

@RayFernando1337 Did you try running on respcode.com to get multi model orchestrating to get the best results.

0 0 0 17 0

View Details

RootFTrust @parallel_beings

5 months ago

@gdb Exactly, that's the reason we built respcode.com to collaborate between models on fixing tests for more accurate results.

0 0 0 17 0

View Details

RootFTrust @parallel_beings

5 months ago

respcode.com/blog/cross-arc…

0 0 3 127 0

View Details

RootFTrust @parallel_beings

5 months ago

@LiorOnAI You should try respcode.com

0 0 0 18 0

View Details

RootFTrust @parallel_beings

4 years ago

@iGio90 I resolved the touch screen issue after Kasan+kcov in my pixel 4xl using this blog reference. blog.senyuuri.info/2021/03/01/fix… Similarly I tried using an alternative method for my Pixel 4XL and it worked. github.com/parallelbeings…