ChatGPT for Embedded Systems - A game changer or Pure hype?
ChatGPT is all the buzz these days with a stream of articles, and worst of all, unending streams of LinkedIn posts. This talk covers my explorations using ChatGPT for embedded systems use-cases, from simple snippets and questions, to a reasonably complex design on a MCU. This talk will uncover the nature of ChatGPT as we put it through its paces to try and answer the question “Are embedded engineers replaceable with AI?”.
The answer(spoiler: No), while unsurprising, leads to some very interesting thought experiments on what the future of this field looks like as we head into a new world of AI-assisted design and learning.
What main reason did the speaker give for why ChatGPT will not replace embedded engineers?
Hi Harsha, great talk! Loved your analysis and thorougher assessment of using chatGPT 3.5 and 4 to assist you to write firmware.
Early days for these LLM's but things will accelerate quickly. Exciting and scary times ahead!
Very interesting talk! Liked the approach of giving your journey clearly of experimenting with chatGPT.
Thank you very much for the kind words
Thank you for sharing the output of your months of evauation! the changes you were able to perceive in the development of the model are striking, so the comparison was valuable. I thank you for spending your frunstration on it so I didn't have to. :-) Still, I am excited to try using it more on personal projects where sending code and ideas are not proprietary to my customer or my company.
One of the best parts about going through the frustration is being able to publicly share the pain. Very cathartic :D
Hope you have great fun in your GPT Adventures Dan
Thanks for the great talk. The method of delivery as a timeline of your experience made all the difference to me. I especially loved the analogy of treating the model as you would teaching another person. I wouldn't have guessed that, with treating it as a tool rather than what it really is.
Thank you for the kind words Goutham. I found the greatest success in 'feeding' it information it should bias itself to (like a code snippet) rather than just open-ended design. Maybe something they can fix/improve at a later point since it can get cumbersome with a pure 'chat-style' interface.
That was really mind opening, thanks a lot for this insight, Sree Harsha!
Glad you enjoyed it rokath. Definitely give it a spin if you're curious; it does help when you're feeling a little bit of lethargy to start a new project.
Thanks for investing your time in this, I've spent a lot of time with the nRF5 SDK so the BLE streamer example was really interesting. I watched someone code a FlappyBird clone in C# with ChatGPT on YouTube recently and saw the glimmer of possibility but this example was much closer to home. :)
Since a lot of companies are restricting use of chatbots due to IP concerns, I'm so intrigued to see how this technology will integrate into the software engineering world...
Somehow seeing ChatGPT spit out embedded code with vendor specific libraries made it more 'real' to me as well.
Your point on IP concerns is very timely and valid. Something which needs to be resolved for this to have any legs for professional development
Interesting talk, your findings align with much of what I have found myself. A key area that concerns me with adopting code written by any third party is the maintainability of the code, personally I use TDD so that I have a good set of regression tests for code that I write. I've been experimenting with chat GPT in C++, asking for code along with unit tests. Some examples have worked really well, others less so, and of course how do I know that the tests are correct? As you mentioned I have found at times that it feels like pair programming. It will be very interesting to see hoiw this impacts our industry over the next few years...
Yep, and I think a few of these may not be a fundamental change in the way we train transformer models like GPT/ChatGPT; it just maybe giving it access to tools with which it can iterate with. i.e., give it full command-line/dev access in a machine it can run compile tests on.
Interesting talk, thanks.
Fascinating talk on a subject I wasn’t familiar with. I’m surprised it doesn’t have access to an online compiler to validate its generated code, before confidently presenting it to you. I tried it myself, asked it if the code would compile, it immediately apologised and made a correction. I enjoyed the unfiltered story.
We're not too far from it. Folks have already started giving GPT access via the API to terminal-like functionality so it can iterate and test code. If you're interested, take a look at AutoGPT(look at the github from Significant-Gravitas). Rapid progress, and we may just be looking at growing pains here.
Thank you for the talk. I've wanted to look into this more but just haven't been able to take the time to experiment. My concern has been the "making stuff up" problem LLM can have with scientific information, but as partner to help set something new up or experiment I really see the value.
Glad you liked it. Part of the reason I kept it unfiltered was so people who didn’t have time to experiment get atleast a rough view l.
Completely agreed on ‘making stuff up’. It still remains to be seen if LLMs can truly eliminate hallucinations. It may just be that systemically these types of models never get away from this and we just live to accept some risk(or figure out other tools to augment them)
Great talk, which coincides with my experience with ChatGPT so far. Especially - as you said - as a "pair programmer" ChatGPT is valuable to point you in the right direction for basic tasks. I even used it to solve math problems, with great solutions. We'll see what the future holds up for us in AI technologies... Maybe there is a talk from your experience in a year or so... Thank you so much!
Hi Thomas,
Thank you for the kind reply. If the rate of progress is anything to go by(and AI hasn’t taken over the world), next year is likely going to be unbelievably different. About time we all got a little personal helper :)
Interesting talk! Thanks for sharing to save us to try!