Always remember that when an LLM prints the beginning of a text, it has no idea what the end will be.
Therefore, when it says "The answer is yes, and this is why:" the text after "why" would most likely be a very elaborate lie combined with gaslighting in case "yes" was the wrong answer.
Okay here's my argument for exactly how I think the "AI uses a bottle of water" statistic looks like it was miscalculated. It's a little in the weeds. Link below
@Mark_Graph That seems like an invocation of the No True Scotsman fallacy to me.
Nonetheless the right measure to chart would be retail dollars spent per KWh consumed at home. I’d be interested to see those data if it existed (which I doubt, since all-home consumption isn’t tracked AFAIK)
It’s been a tough stretch with GitHub availability.
The feedback isn’t wrong. People depend on us, and we haven’t met that bar.
Just know there are humans on the other side of this, reading all your comments and working as hard as we can to improve it. https://t.co/tBfKUrbdRf
The 2-hour marathon was this generation's 4-minute mile. People debated whether human physiology could do it. The once-inconceivable achievement was broken today not by one runner, but two! Sabastian Sawe just ran 1:59:30 in London.
Four photos of Wimbledon’s Centre Court show how tennis has shifted: from grass worn across the whole court (serve-and-volley era) to damage almost only at the baseline today. A slower, power-driven game played from the back.
The 1970 picture (top left) shows the grass played to a pulp all over the court, from baseline to net. In 1980 (top right), turf damage is visible in a T-shaped zone beyond the baseline. This damage to “volley valley” largely evaporates by 2005 (bottom left). By 2023 (bottom right), the service boxes appear entirely untouched, with only the baseline area showing wear and tear.
Source: https://t.co/hnkzeS9Q6N
Excited to announce the new preview for Microsoft Foundry Agents 🎉! You can now build, run, and deploy your agent using any model, any framework, any harness in the cloud 🧑💻 - check out the demo below
This is not just any cloud compute environment; it's an agent-optimized platform with:
🖥️ Persistent microVMs - securely scale up and down without losing context
🛠️ Built-in tools (1000+)
👀 Observability and evaluations
👷 Guardrails
🔐 Private networking... and more
Saw the EML paper blowing up yesterday so I thought: “What if I build a full LLM inference engine with it?”
So I did.
Meet emilio — an end-to-end inference engine where every single multiplication is replaced by the EML primitive:
eml(x, y) = exp(x) - ln(y)
No *. No +. Just this one wild operator (plus the constant 1) powering all the math: linear layers, attention, everything.
It runs Qwen2.5-0.5B-Instruct (494M params) completely in EML and actually generates coherent text.
→ ~5.5 tok/s on CPU
→ ~30 tok/s on Apple Silicon GPU
Not competing with llama.cpp on speed. This is pure proof-of-concept: showing the math actually works at transformer scale.
A celebration of mathematical minimalism in the LLM era.
Repo: https://t.co/NWN9w1Yz2F
My paper (how I built + optimized it): https://t.co/parp7vkAwj
If you like weird math, extreme minimalism, or just watching someone go “hold my beer” after reading a paper… star it ✨
#LLM #MachineLearning #Math #Rust