> Multiple studies, including from organizations notable for their bullishness on AI, suggest LLMs are not accelerating productivity or otherwise generating positive returns on investment for most businesses.
Maybe this is true for most businesses, but that's because most businesses are not really leaning into agentic AI yet. It seems software devs started to use it in 2025 and OpenAI and Anthropic have tried expanding the use cases with their GUIs and webbased agentic systems, which I think means there is incredible growrth opportunity when the masses start adopting it.
I would not think in terms of "how much" context, but really just how the harness manages it. I've been fiddling around with harnesses earlier this year and the rabbit hole goes very deep. I eventually had to drop it because it consumed me, and I can only do one thing well. But what I learned is that there is a lot on the table at the harness level.
The elephant in the room and the constant tension is that the models do get "used" to a certain harness (trained for it by the model labs), but they also generalize well. So one really advanced mechanism could fall short because the model could fail to generalize, whereas in other cases the model might handle large deviations from how a conventional harness works better than expected. This is where the gains are.
I wonder how feasible decompilation is. The firmware is used in many boards. An open, cleanroom reimplementation (derriving hardware details from the firmware, then reimplementing cleanly in embedded C++ or rust) could fix the issues, that would be very interesting.
That's true. It's a two way street. I feel I gain more from what the model offers me, than what I provide back. Of course, local would be a thousand times better. One day I'll have 20k to spare and get the hardware.
Yes, so different SoC, same WiFi, same issues, isolating the issues to the WiFi chip (common denominator). Carrier was also ruled out since I tried a custom and Raspberry offical one.
It's the same chip, renamed (which is silly) after the Broadcom IoT division was bought by Cypress, which is bought by Infineon. So technically, it's the Infineon CYW43455 (formerly Broadcom BCM43455).
I tested te Pine64 SQQuartz CM4 compatible SoM, but it uses the same Cypress CYW43455 WiFi as the CM4/Pi4, which crashes on me under load (not the kernel, the chip itself). There is no fix or easy mitigation, other than lowering the load which only reduces occurence. I't odd that not many are talking about this.
The CM4 and Pi 4 use the Cypress CYW43455 for Wi-FI, which I found to be unreliable. Under load, it crashes, and it can't be fixed in the kernel.
Frustratingly, many other SoMs and SBCs use this chip. I don't understand why there is no recall on this - in my opinion - trash product. It's an odd feeling to be the only one experiencing this. I am 100% confident this is a design flaw in the CYW43455. I've seen this also on a Pine64 CM4 compatible SoM, which has a completely different SoC. This is on both the official and custom carrier. Cypress, fix your shit!
I used to care about this initially, but now I don't. How can them having my data, mostly engineering work on niche (small market) products hurt me? An they need to know what's going on, how else would they ever improve the product? Benchmarks?
What could happen, even in countries with strong socialist worker protections (western europe) or high cost of living (US, europe, many others), is that labour will become cheaper again. We'll be "poorer", less able to import, but on the other hand may be able to do more locally. We may also see larger economic class differences, as you essentially said youself.
In a way this could be a downgrade, we'll be less incentivived to optimized or automate, and will be able to just throw bodies at problems.
From reading xpct's comment in this thread it seems like they use effectively a very deep and wide pipeline by tiling the cores without central clocking or dispatch. In theory if a GPU is centrally organized it could simulate this behaviour, but I imagine that creates transfer overhead. Just speculating though.
They're very afraid of the competition, this really fits their modus operandi. They don't want to reveal inner workings any more than needed for the product to be usable.
Thankfully anybody can spin up a VM and vibe up something similar. It's not like the Windows moat or anything, it's not sticky.
Not using assistants, but I had it do quarterly reporting. Gave it the password to my webbased accounting software which does not have an API. Told it to write a CLI for itself to access the administration. I only did the bank exports to CSV. Told it to look in my Thunderbird mailbox to find invoices and let it do the quarterly reporting. Main issue: Many foreign vendors don't use invoices, which means my Dutch, invoice-centric accounting software kinda sucks for international import-heavy operations.
I have it use `playwright-cli` which runs a headless Chrome session, with relatively structural representations of websites without the HTML bloat, plus image viewing options. The builin web fetch tool indeed seems blocked. Headless Chrome is just Chrome, you can't block it. I think even the "Are you human" checkboxes are okay, since LLMs are good at coordinate guessing and initiating clicks.
I've been doing PCB layout with Opus 5.5. Maybe this was possible easlier and with Chinese models, but I feel this really pushes the model. These are 10hr jobs, so also requires good stability over time. I suppose this could be tested. As the models get better, I also get better at using them. So I'd say yes, we do more advanced things with the models every year.
Every time I tried Flatpak it did not work, always some other problem. Maybe this got fixed nowadays. I did not find it user friendly either, but this could be addressed by the distro.
AppImage actually works and is simple, so I use that.
Snap is trash because autoupdate cannot be disabled without code changes (what?!) last time I checked.
Why this model split? And are your work volumes so high that you prefer these cheaper models over a subscription or are you simply trying to avoid the 200/m? (which will only go up).
In my experience the MCP is not needed, it just runs headless FreeCAD using CLI. The MCP can be useful if you want to run FreeCAD on a VPS, and then use ChatGPT e.g. on your phone (optionally via voice). But for local I'd go with the CLI. But they'll do just fine with the MCP, which probably wraps the Python API anyway (ideally, otherwise it could lack some features).
Astra in Codex CLI is good with FreeCAD if you let it use the FreeCAD Python runtime in a headless session. It is good at relying purely on natural language to design and revise. To follow specific CAD techniques you need additional prompting. Highly recommend. Uses quite a lot of tokens.
I've been using Droid Sans Mono for everything mono for a long time and can recommend it. But this font might be great too. Any Droid Sans Mono lovers out there?
For people who prefer a touchscreen input over a laptop, and if they'd want to do agentic work, or perhaps text to speech, image generation, music production, fancy spellcheck while traveling?
I have to be careful with hyperboles, seems like my comment got flagged. I meant "the things I cannot find with Google I can find with Yandex". I actually have no idea about overall censorship, so "zero censorship" is a bit insane of me to proclaim in retrospect.
I guess what I meant was "I can always find what I want", I have not tried searching for things that the Russian regime may find sensitive. So the better term would be "censored very differently", so it fills up Google's blind spots nicely.
Ukraine has suffered much more material damage and loss of large swaths of land, and may have lost more civilians. Russia did lose a lot of troops, but their population is also larger. Russia did experience a brain drain, but Ukraine as well. I find "beating Russia" an odd take.
> does still cover marginal inference cost
Simply comparing to the larger models on OpenRouter implies that the pure hosting costs (equipment + power + minimal overhead) still exceed plan pricing if we assume all users always use their full weekly allowance.
So my conclusion is that it only works because the majority of users doen't fully utilize their allowance. Last month I used almost nothing of my Claude 20x plan (did use Codex though).