Would you be willing to elaborate on the ways in which the $200/month subscription is better than the $20/month subscription? I’m genuinely curious - not averse to paying more, depending on the value.
With Pro-Mode, I can paste an entire firmware file—thousands of lines of code—along with detailed hardware references, and the model can actually process it.
This isn’t possible with the smaller context window on the $20 tier. On the Pro plan, I’ve pasted like 30+ pages of MCU datasheet information plus multiple header files in a single go. The model is then reasonably capable to provide accurate, bit-twiddled code, many times on the first try. Is it always working on the first go? Sure sometimes, but often there's still debugging, and I don't expect people that haven't actually tried to do it before without AI could do it effectively. However, I can do a code diff using tools like beyond compare (necessary for this workflow) to find bugs and/or explain what happened to pro-mode perhaps with some top level nudge for a strategy to fix it, and generally 2-3 tries later we've made progress.
2. Deeper understanding, real solutions
When I describe a complex hardware/software setup—like the power optimization for the product which is a LiPo-rechargeable fan/flashlight, the Pro-Mode model can understand the entire system better and synthesize troubleshooting approaches into a near-finished solution, with 95–100% usable results.
By contrast, the non-pro plan can give good suggestions in smaller chunks, but it can’t grasp the entire system context due to its limited memory.
3. Practical Engineering Impact
I’m working on essentially the fourth generation of a LiPo-battery hardware product. Since upgrading, the Pro-Mode model helped us pinpoint power issues and cut standby battery drain from 20 days to over a year.
Like, this week it guided me to discover a stealth 800 µA draw from the fan itself when the device was supposed to be in deep sleep. We were consuming ~1000 µA of power when it should be about ~200 µA. Finally, discovered the fan issue and achieved 190 µA without it in the system, so now we have a move forward to add a load switch so the MCU can isolate it from the system before it sleeps. Bingo we just went from a dead battery in ~70 days (we'd already cut it from 20 days to 70 days with firmware changes alone) to now it should take about 1 year for it to drain. This is the difference between end users having zero charge when the open the box to being able to use the product immediately.
4. Value vs. Traditional Consulting
I’ve hired $20K short-term consultants who didn’t deliver half the insights I’ve gotten in a single subscription month. It might sound like an overstatement, but Pro-Mode has been the best $200 I’ve spent—especially given how quickly it has helped resolve engineering hurdles.
In short: Probably the biggest advantage is the vastly higher context window, which allows the model to handle large, interrelated hardware/software details all at once. If you work on complex firmware or detailed electronics designs, Pro-Mode can feel like an invaluable engineering partner.
How much context do you get on the $20 plan? I run llama3 at home which technically does 128k but that eats vram like crazy so I can't go further than 80k before I fill it (and that is with the KV store already quantified to 8 bit).
I've been thinking of using another service for bigger contexts. But this may not make sense then.
My 2 cents on the long context (haven't used Pro mode, but older long context models):
- With a statically typed language and a compiler, it's quite easy to automatically assemble a meaningful context with 1-2 nested calls of recursive 'Go To Definition' and including the source from that. You can use various heuristics (either from compile time or runtime). It's quite easy to implement, we've done this for older, non-AI stuff a while ago, for trying to figure out the impact of code changes. If you have a compiler running, I'm pretty sure you could do this in a couple days. This makes the long context not super necessary.
- In my experience, long context models can't really use their contexts that well. They were trained to do well on 'needle-in-the-haystack' benchmarks, that is, to retrieve information that might be scattered anywhere in the context, which might be good enough here, but asking complex questions that require the understanding the entire context trips the models up. I tried some fiction writing with long context models, and I often found that they forgot things and messed up cause and effect. Not sure if this applies to current state of the art models, but I bet it does, since sequencing and theory-of-mind (it's established in the story that Alice is the killer, but Bob doesn't know that at that point, models often mess this up and assume he does) are still active research topics, and current models kinda suck at it.
For writing fiction, I found that the sliding window of short-context models was much better, with long-context ones often bringing up irrelevant details, and ignoring newer, more relevant ones.
Again, not sure how this affects the business of writing firmware code, but limitations do exist.
I don't have the pro plan, so can anyone compare it to the results from the new Google models with huge context windows (available in aistudio)? I was playing around with them and they were able to consume some medium (even large by some standards) code bases completely and offer me diffs for changes I wanted to implement - not the most successful ones but good attempts.
"Like, this week it guided me to discover a stealth 800 µA draw from the fan itself when the device was supposed to be in deep sleep."
Was this context across a single datasheet or was Pro-Mode able to deduce from how multiple parts were connected/programmed? Did it identify the problem, or just suggest where to look?
How do you input/upload an engineering schematic or cad file into chatgpt pro-mode? Even with a higher context window, how does the context of your project get into chatgpt?
When I get stuck or have a larger task or refactor, I'll paste in multiple files. So at the $20/mo you get rate limited pretty quick. I made a tool to easily copy files https://pypi.org/project/ggrab/
Have you tried using Cursor? I’m using it with Claude models but it works with ChatGPT ones too. It’s a fork of VSCode with an AI chat sidebar and you can easily include multiple files from the codebase you have open.
Not sure if it’d work for your workflow, but it’s really nice if it does.
I was looking at the team $25/mo last week and it had mentioned priority access but that language is gone and instead I see Team data excluded from training by default. It seemed worth the difference, but now less clear with changes in description. Basically I just want to know if it's a 'superset' better or has tradeoffs.
bobjordan|1 year ago
1. Larger Context Window (128K)
With Pro-Mode, I can paste an entire firmware file—thousands of lines of code—along with detailed hardware references, and the model can actually process it. This isn’t possible with the smaller context window on the $20 tier. On the Pro plan, I’ve pasted like 30+ pages of MCU datasheet information plus multiple header files in a single go. The model is then reasonably capable to provide accurate, bit-twiddled code, many times on the first try. Is it always working on the first go? Sure sometimes, but often there's still debugging, and I don't expect people that haven't actually tried to do it before without AI could do it effectively. However, I can do a code diff using tools like beyond compare (necessary for this workflow) to find bugs and/or explain what happened to pro-mode perhaps with some top level nudge for a strategy to fix it, and generally 2-3 tries later we've made progress.
2. Deeper understanding, real solutions
When I describe a complex hardware/software setup—like the power optimization for the product which is a LiPo-rechargeable fan/flashlight, the Pro-Mode model can understand the entire system better and synthesize troubleshooting approaches into a near-finished solution, with 95–100% usable results. By contrast, the non-pro plan can give good suggestions in smaller chunks, but it can’t grasp the entire system context due to its limited memory.
3. Practical Engineering Impact
I’m working on essentially the fourth generation of a LiPo-battery hardware product. Since upgrading, the Pro-Mode model helped us pinpoint power issues and cut standby battery drain from 20 days to over a year. Like, this week it guided me to discover a stealth 800 µA draw from the fan itself when the device was supposed to be in deep sleep. We were consuming ~1000 µA of power when it should be about ~200 µA. Finally, discovered the fan issue and achieved 190 µA without it in the system, so now we have a move forward to add a load switch so the MCU can isolate it from the system before it sleeps. Bingo we just went from a dead battery in ~70 days (we'd already cut it from 20 days to 70 days with firmware changes alone) to now it should take about 1 year for it to drain. This is the difference between end users having zero charge when the open the box to being able to use the product immediately.
4. Value vs. Traditional Consulting
I’ve hired $20K short-term consultants who didn’t deliver half the insights I’ve gotten in a single subscription month. It might sound like an overstatement, but Pro-Mode has been the best $200 I’ve spent—especially given how quickly it has helped resolve engineering hurdles.
In short: Probably the biggest advantage is the vastly higher context window, which allows the model to handle large, interrelated hardware/software details all at once. If you work on complex firmware or detailed electronics designs, Pro-Mode can feel like an invaluable engineering partner.
wkat4242|1 year ago
I've been thinking of using another service for bigger contexts. But this may not make sense then.
torginus|1 year ago
- With a statically typed language and a compiler, it's quite easy to automatically assemble a meaningful context with 1-2 nested calls of recursive 'Go To Definition' and including the source from that. You can use various heuristics (either from compile time or runtime). It's quite easy to implement, we've done this for older, non-AI stuff a while ago, for trying to figure out the impact of code changes. If you have a compiler running, I'm pretty sure you could do this in a couple days. This makes the long context not super necessary.
- In my experience, long context models can't really use their contexts that well. They were trained to do well on 'needle-in-the-haystack' benchmarks, that is, to retrieve information that might be scattered anywhere in the context, which might be good enough here, but asking complex questions that require the understanding the entire context trips the models up. I tried some fiction writing with long context models, and I often found that they forgot things and messed up cause and effect. Not sure if this applies to current state of the art models, but I bet it does, since sequencing and theory-of-mind (it's established in the story that Alice is the killer, but Bob doesn't know that at that point, models often mess this up and assume he does) are still active research topics, and current models kinda suck at it.
For writing fiction, I found that the sliding window of short-context models was much better, with long-context ones often bringing up irrelevant details, and ignoring newer, more relevant ones.
Again, not sure how this affects the business of writing firmware code, but limitations do exist.
egeozcan|1 year ago
savorypiano|1 year ago
Was this context across a single datasheet or was Pro-Mode able to deduce from how multiple parts were connected/programmed? Did it identify the problem, or just suggest where to look?
welder|1 year ago
uncomplexity_|1 year ago
scrollaway|1 year ago
keizo|1 year ago
mcintyre1994|1 year ago
Not sure if it’d work for your workflow, but it’s really nice if it does.
franze|1 year ago
No worries that you run put of prompts for o1. which allows for more experimentation and creativity.
karmakaze|1 year ago