chris_money202

So your proposal is for Microsoft to invest their free cash in helping support a bunch of AI coders maintain their pet projects? Sounds like you should be a CEO!

Intel started making DRAM in 1970, Micron was founded in 1978.

Nvidia buys the memory it uses on its GPUs, same as all other ASICs.

To give some context, Intel started making DRAM, I think they were actually the company that came up with modern memory techniques. They exited the market and pursued a more lucrative moat with CPUs.

RAM chips are not hard to produce compared to many other types of semiconductors; Intel started in the memory game and left because the margins weren't great and they were going to fold. The failure rates on these chips are actually very tolerable; you can have a very bad yield and still have a viable chip due to things like ECC.

In the short and medium term, it probably won't be more economical to produce for OpenAI. Where OpenAI is benefitting from their own chip is being able to tailor it to their models and workloads. When you buy off the shelf Nvidia, its not perfectly tailored and OpenAI has to spend marginally more to run off that chip. At the scale OpenAI is operating at and plans to operate at, that margin becomes pretty big $$

Opening large PRs is bad engineering practice, regardless of if an LLM or human wrote the code.

Would recommend your senior team look into stacked PRs, which was made to handle this scenario exactly.

Writing code and being formally taught a coding languages syntax and toolchain are quite different.

Most computer science programs don't teach "coding" though. There is pretty heavy emphasis on datastructures, algorithms, and system design. I don't think AI atrophies those skills / knowledge as much as the coding language and I think in many cases actually improves them, especially if you are reading the code that is produced and at least understanding the flow of it.

I agree that AI reliance is hurting engineer's language understanding, but I don't know how much we should care.

Anyone can train LLMs on public GitHub repos

I think GH/Microsoft got caught by surprise to the huge usage number increase as a result of LLMs. Microsoft started rolling all of its DCs to OpenAI around 2019-2020 to begin LLM training, then we saw a huge GH increase in late 2024 and Azure was just out of capacity for infinite GH scale.

Azure is at capacity; GH offers a lot of services for free or heavily discounted. Especially when its public repo. These aren't profit generating for Microsoft, so they are not the priority with the capacity Azure has available. GH is probably struggling with these capacity and budget constraints. They also still have some AWS and non-azure DCs in use.

Once Azure's massive spend is online things will slowly get better.

It was probably written 95% by an LLM and has only a few people (if any at all) that understand major parts of the system end to end.

It will eventually have the same uptime issues as GH (or worse) with less ability to provide stable fixes.

GitHub is crashing due to scaling pressure. Something GH is much more equipped to solve long term.

The kv cache is wiped as soon as you get your answer, cloud hosts are not going to hold the GPU memory for your entire session. You're probably referring to some agent level cache

I don't know if that's the whole point, but I agree with the sentiment, why would a startup be working in legacy code and where would that code come from if this is truly the start of something.

OP might just be working at a small software company or for one that broke from a bigger one and is now "startup" like?

My thoughts on this is there is a deliberate act going on that is doing more for retention than just the simple act of using a (fountain) pen.

If I am in a lecture and I am writing down everything I hear as fast as possible in one ear, on paper, out the other style I don't think writing is any different than typing the same way.

If I am in lecture and I am trying to stay deliberately between the lines of paper, then I have to take the info in, process it in a way where I analyze and summarize it, and then put it on paper neatly. I have to do that for brevity and to keep up with the lecture. Its not the act of just writing, its the analyzation that occurs before the words hit the paper. This can be achieved with typing as well if you deliberately style your notes.

I think sometimes we correlate this in too simple of terms to use pen = better, the pen forces the deliberate action of analysis easier than typing does.

oh great now I have to worry about the wing tips flying off midflight as well

Right, a heat exchange is reusable though on earth and can be recycled across multiple SKUs and thus depreciates longer. In space similar to how you can’t repair you also can’t swap.

Repairability and depreciation are the main problems. A earth data center can be repaired, depreciated, and recycled at EoL recovering some of the costs. SpaceX datacenters are a total write off from the moment they are launched.

Based on description it doesn't sound like SoC design, it sounds like a book about RISC-V microprocessors. To the untrained eye, that sounds similar, but a microprocessor is one part of an SoC, and sometimes it can be a very small part based off the role the microprocessor plays.

You don’t need to serve raw html, you can just open it in a hypertext markup language viewer (I.e a browser). Just send people the index.html

Don't be fooled, there is politics, opinions, and less rigor in science as well.

"How it Works" -> "Idk" OP probably

Wow that’s like such an obvious thing that makes so much sense. I have no doubt something like “Hands on, Hands on, Hands on” will start chirping in near future

I think future is probably more similar to speculative execution (inference/decoding). A small LLM is used to speculate and a large LLM is used to confirm if needed. If the small LLM is accurate enough on N tokens it’s cheap for the large LLM to say looks good and keep moving along.

My guess is area becomes problematic. Chip becomes bigger, harder to manufacture without defects, costs rise as a result of those defects.

Bonus though, Apple is using "defective" chips in other products now such as the Neo which will help some with costs, but overall, you can't just add memory to the chip because the chip can become very yield sensitive, the process has to be there to produce the yield effectively.

I still don't know if I am sold with that. .NET is also a runtime so how do you handle the user story? If your users are also on Windows Lite then they have to manage .NET version or you have to package .NET with whatever you build. If your users are on full Windows, wouldn't it just make sense for you to build in same environment as your Users? Especially since IT would have to manage two separate operating systems if devs went Windows Lite and say Sales using the target app was on Windows.

This whole thing makes sense for indie devs or build VMs but breaks down for Enterprise pretty quick, and Microsoft is much more friendly to Enterprise customers than indie devs.

OP says Windows needs builders to make applications and that is what Windows Lite would enable (or expand) but then says Windows Lite shouldn't include .NET one of the primary frameworks to build applications on Windows.

Or hear me out, we improve life here on Earth...

If you’re truly serious about local AI you’re not using a Mac. A Mac is for people who want the best of both worlds, GPUs are much faster if memory is equal.