Every second of free time you have ever had in your life was subsidized by the existence of the combine harvester.
09.03.2026 19:53
👍 153
🔁 19
💬 6
📌 2
Every second of free time you have ever had in your life was subsidized by the existence of the combine harvester.
What's interesting about basically ~all possible monetization is that it only applies to people using bsky's PDS and/or client. I think charging for PDS space makes a lot of sense actually.
I mean, she's still here, just in a different role
96gb 5090 waking up in hospital bed "where is my wife 5090" "who do you think gave you the memory"
FrankenGPU donor 😔
When my car got totaled I almost got a used EV, but unfortunately as a rentoid I wouldn't be able to charge at home and DC fast charging is kinda annoying, (more) expensive (than charging at a slow home charger), and hurts the batteries if done a lot :(
perhaps 'the market' acts like a moron (believes politicians) because every trader expects every other trader to be a moron due to (occasional) prior behavior, thus behaving like a moron because that is how the market moves, due to the traders thinking others to be morons
The new role seems (fine/good), but I'm a bit concerned about a VC person being CEO, even briefly...
Early access customers have seen similar patterns. On a ZFS encryption refactor in TrueNAS's open-source middleware, Code Review surfaced a pre-existing bug in adjacent code: a type mismatch that was silently wiping the encryption key cache on every sync. It was a latent issue in code the PR happened to touch, the kind of thing a human reviewer scanning the changeset wouldn't immediately go looking for.
A future where every line of code has undergone so much (AI) review that bugs are seen as rare. Just like median-artisan goods vs mass manufactured, people end up preferring the latter because it's better than the worst-through-medium human made stuff and cheaper than stuff from the best humans
To me, it seems more and more like the only way to avoid getting eaten by a biglab is to develop something they will have no interest in copying; something unique, odd, or highly unlike existing tools. Deeper integration than 'the AI does this job now'
Only for team/enterprise for now, but it looks like Anthropic has more or less just killed all of the AI code review wrapper startups. This seems to be super in-depth (expensive!) and supposedly less than 1% of reviews are flagged incorrect. Excited to see if pro/max gets this.
it is actually really hard to write post text for these. i have already written 5000 words. i have a title, a subtitle, a header image and an opening. what do you write to add to that
anyway all that time talking about philosophy of mind might pay out now?
Dear god, they chudified the Marathon style
GLPs are cool because I don't eat much but then when I do eat it makes me want to throw up immediately :/
It still has a layer of 365ium slop over top of it, helps get it to the Microsoft guaranteed 8gb of memory consumed requirement
Breaking news: Microsoft creates absolutely dogshit clone of Claude cowork
Like the other replies say: stuck for network/infra reasons. The reason new models seem to 'break' mid-thinking is probably due to interleaved thinking, which lets the model drop in and out of thinking mode, and each phase is probably handled as a separate request
Real breaking news there. Thanks google
In my experience this matters a lot for how good the code ends up... Claudes generally seem to like working on AI related projects especially and tend to do much better on those
The most truth-seeking AI
Oil & LNG prices explode -> electricity costs surge -> even less DC hardware can be deployed near term -> Gulf states reduce funding for local DC projects and VC funds to AI corps due to local attacks/infra damage -> Push timelines out by at least a few months
Is Trump a pauseAI plant?!
New post! Solar storms are damaging and expensive, are a tail risk for catastrophic harm, and can be averted straightforwardly and cheaply (only we haven't done so).
www.lesswrong.com/posts/ghq9Ew...
It's a weighted average of each possible token's embeddings, weighed by probability
System prompt: 20000 tokens
Skill definition: 4000 tokens
Output style with full explanation of both my fursona and the fursona the agent is supposed to communicate as: 110000 tokens
Hooks: 2000 tokens
Some who is good at the economy please help me budget this
my context window is dying
alright, it hasn’t crashed and burned yet so, try my dumb game?
canyoupasstheturingtest.com
This is pretty similar to a thing I made once, though that was focused on guessing what specific model generated some text. My best on this so far has been 70-80%. The encyclopedia questions are brutal. ChatGPT likes to use () a lot, but that's about the only tell I can see on those usually
That's correct yes, it's showing the logprobs. For mix tokens, it takes those and creates a weighted average - that is what the model actually ends up getting as the 'token', an interpolation between all the options based on what the model thought was most likely
OBS requires more than 20 megabytes of VRAM to do it's thing, so unfortunately I had to drop down to a 2B parameter Qwen to show this off, but here is what it looks like! Also, here is the github repo if anyone would like to try it themselves! github.com/Isolyth/unce...
I find that odd too. I'm very skeptical of static maps working for more complex brains as well. Also can't say I would want to be implemented on a system with only 95% accuracy...
Essentially. Everywhere.