The Nazi soldiers tried hard and were just following orders too. Look where it got them 😀
07.03.2026 00:33
👍 0
🔁 0
💬 0
📌 0
The Nazi soldiers tried hard and were just following orders too. Look where it got them 😀
Don't need low latency for inference, just efficient training.
Everyone using Claude and chatgpt is 15-100ms away from the services they use.
Slightly related but I noticed over the past few weeks that the Starbucks app only seems to ask how your ordering experience was if the wait time is less than 10 minutes.
Like they are trying to reinforce your good vibes so when their service sucks you don't feel quite as negative about it.
Claims that Qwen 3.5 are even close in ability to any recent Sonnet model are highly exaggerated - at least for the unquantized 122b model I've been testing (writing bash scripts)
Very friendly though.
If random video game loot boxes are illegal - what about Pokemon cards or trading cards in general?
Paying a fixed price for a pack in which you might find something valuable. Zero difference to counterstrike loot boxes.
This is a slippery slope.
www.joneswalker.com/en/insights/...
You can't round a number and then recover the decimals later by looking at it from a different angle.
That's a very charitable interpretation, which doesn't align with the claim that their resulting system exceeds the original precision baseline.
The best you can do is recognize that the model is in a region where it's lost fidelity and gracefully degrade to an "I don't know" rather than letting it confabulate with confidence.
If quantization has destroyed the information, no amount of geometric steering is going to recover it. The bits are gone. You can't tether your way back to knowledge that's been truncated out of the weight matrix.
Utter bullshit.
Sales pitch disguised as research.
For this to work, the author would have had to discover some way of identifying what is true in the representation space - which, if they had done, would be a bigger deal than this paper.
It feels like a charade to swap out Anthropic for OpenAI.
On a totally unrelated note,
I wonder which part of the Whitehouse Ballroom Greg Brockmans donation is paying for.
Wow. I guess those Whitehouse "Ballroom" contributions from OpenAI Greg Brockman really come in handy when negotiating federal contracts.
I just had to make this.
Access to AI is going to become such an important part of life that employers should start adding AI to their compensation packages. (and I predict they will)
Would you choose Claude or Chat GPT during open enrollment?
Great distraction from the real news which is the suspected theft of 1B customer records from Salesforce.
www.sfgate.com/bayarea/arti...
The most valuable data AI companies have now is user inputs - Estimated at 2.5B per day for OpenAI.
In addition to training AI on a huge corpus of knowledge, they'll teach them to predict what the user will ask next.
One day we'll use their service and it will give us a response with no prompt 😀
50 from the five tens
6 ( the 1 and 5 in the 11 and 15)
6 ( the 2 and 4 in the 12 and 14)
3
50 + 6 + 6 + 3
50 + 10 + 2 + 3
65
I feel this weird change in the world like there's a kind of conscious awakening happening. Connecting with new people and sharing ideas.
It's really quite refreshing amongst all the other bad stuff going on.
I wiped and sent my echo dots for recycling a few weeks ago. I think a custom local home voice assistant is the way to go from now on.
Then you give them the price and the person with the idea says "that's too expensive to write an app - anybody can write apps"
Anybody eh? 🤣
Organization controlling current president raises alarm bells that previous president was also controlled by other people.
#nottheinion
www.foxnews.com/politics/bid...
Anyone hacked Claude code yet to make it work with local models?
If not I know a guy,Claude, who'll do it for me.
Two comments on Claude Code.
1. It is very good
2. It eats tokens for lunch and is prohibitively expensive for the kinds of tasks you would need it for.
What are the constraints?
Built in knowledge only, or can I pass it the EPS spec?
Can I use agents with various roles working together, or only a single instance of a model?
What have you tried? This doesn't seem like a huge challenge.
Oh yes that was great.
To imagine your friend maybe or maybe not doing.
I can imagine putting mixtures like that in metal tubes made from aluminum foil and lighting them with slow burn fuses. sometimes with sulphur and charcoal in as well.
Must have been fun times I never ever participated in.
Got into the Claude code agent preview. Either scarcity wasn't a real thing or they liked my python and a nice message I left.
Either way - 😍
Anarchists/ Jolly Roger cookbook has entered the chat... 😀
An AI family wouldn't need food...
100-40 = 60
Add the 4 (104-100)
Add the extra 2 (40-38)
60+4+2 = 66
But to clarify I don't think I either count up or subtract - I just think of the gap between the two numbers followed by the small adjustment.