which GPU?

R9700. I walked into Microcenter on Hallow and picked one up for $1900.

I was also using a 7900 XTX before that which was reasonably capable for single user usage, but I share my server with a friend, and the 9700 is a much better fit. Also nice for doing sub agent workflows that I couldn't really do before.

I think this is similar to how small changes in phrasing can completely change the answer you get from an LLM. This all goes to confirm that LLMs are, at their core, still text completion engines, they (frequently / always?) don't have a firm semantic grip on the topic at hand.

Here's an interesting reference on this that gives a bit more background information:

-- 8< --

Because LLMs (and many other ML systems) are chaotic, it is possible to manipulate them into doing something unexpected through a small, apparently innocuous change to their input. These changes can be illegible to human observers, which makes them harder to detect and prevent.

For example, flipping a single pixel in an image can make computer vision systems misclassify images. You can replace words with synonyms to make LLMs give the wrong answer, or introduce misspellings or homoglyphs. You can provide strings that are tokenized differently, causing the LLM to do something malicious. You can publish poisoned web pages and wait for an LLM maker to use them for training. Or sneak invisible Unicode characters into open-source repositories or social media profiles.

-- >8 --

Source: https://aphyr.com/posts/412-the-future-of-everything-is-lies...

do you have a write up of this?

why does it matter for a word doc. At the end of the day, what matters is the doc file you save. that's your artifact. Weather you use a cli tool, agent or ms word to create that file doesn't matter as long as the content is correct and the person who receives the file can open it.

> What do I do if something happens to you and another lawyer or paralegal has to step in?

This is one of the advantages of running free software. The data formats are open, so all of the data should be able to be imported any other system.

This is one of the advantages of using MS Office. The other lawyer can just use their own copy or subscribe online and not bother with replicating a custom environment.

> $12.5 billion to fix the aging US air traffic control system and Transportation Secretary Sean Duffy wants another $30 billion for additional upgrades as the FAA deals with numerous technology issues.

Why does it cost this much? wouldn't any modern consumer grade computer be able to handle these tasks?

And when the modern consumer grade computer fails in the middle of a busy shift?

buy two? three? redundancy makes for good reliability

> But Anthropic's business practices are scary. They have a reputation for cancelling accounts for random reasons, offer rather bad terms, and have no customer support. I have a strong aversion to services with a reputation for cancelling services arbitrarily. They will apparently cancel based on their system's guesses as to what you're working on. Sometimes they guess wrong. What I'm doing (metaverse clients) is so far out of the mainstream they may guess wrong.

Then use open alternatives as much as possible. Something like Opencode with OpenRouter with Deepseek 4.1 from ZDR provider works very well and costs like 1/20th.

It costs 1/20th the price at token pricing, but an Ultra subscription is still amazing value.

I also think OAI is much better than Anthropic in this regard. They don't try to gatekeep nearly as much.

But I agree that open alternatives are sort of a theoretical ideal. Just a lot harder to get those mega-subsidized.

In practice all of these open models are more expensive than a Claude subscription.

Are they? (genuine question, i've got no experience with Claude) This month I used 1 billion tokens (including cache) with DS4.1 flash and it cost me 7% of my $10 opencode go sub.

If you meant in developer time, then I don't know but it seems very fast to me (rarely thinks more than ~10sec, never more than ~30sec) and usually oneshots what I prompt it to do. I use opencode tui as harness.

In my experience yes. Been using Claude since day one and wanted to experiment with Pi for fun, and try to get my costs down. I found the open models more expensive, eventually had to plug my Pi set up into an OpenAI subscription, that was the only way to get good value… Then I realised OpenAI is way behind Opus 5.5 and their subscription is far less generous with usage, so I’ve come full circle and am back on Claude Code.

i found they were until recently. mimo 2.6 flash and glm 5.3 flash seem to finally be cheaper. i find ds4.1 flash still on the high side.

I found GLM Flash competitive price wise but then my tasks took twice as long to complete. Even if it was cheaper (in my experience it still wasn’t) you then lose out on speed because of the insane amounts of inefficient reasoning.

These days I mostly ask my coding agent to test the APIs. It can use curl create the request, attach all the cookies, whatever is necessary, and send the request, look at the response to all the testing I need. And if I need to manually test something, I ask it to give me a curl command that I can run.

Hey, cool project. Whenever I've seen a deal on a pre-built PC I've always wondered, hey, how much would this cost if I built it myself from parts.

Some feedback on your website. I would suggest moving the "try it" from the bottom of the page towards the top of the page so that visitors don't have to look for it.

Cool beans! Thanks for your comment and pointers. Ill move it down a bit to make it easier to "try"... And btw: please, try it yourself and keep on checking deals! Let me know what you think. The price engine/scoring will change most likely as this is just a prototype that I want to have discussion about. And I hope the service will be helpful! Lets not get ripped off! I check prices every day at around 8:00am

> At its GitHub Universe conference on Wednesday, the company announced that it is now seeing 7.4 billion commits per month. To put that into perspective, developers pushed 65 million monthly commits to the platform in 2024.

Of course, this is because AI is pushing commits without anyone reviewing them.But I think the solution here might be for AI to commit locally and then consolidate a whole feature into a single commit before pushing.