ivanhoe

Bullshit is a tax on society

@LottoLabs $3k is an insane bargain in terms of capability. Consider that people spent $3k on family computers in the 80s and 90s unadjusted for inflation. Our toasters have more computing power now.

Reminder that for like 3k you can give your kids access to models way beyond their need for k-12 and you get no lock in, no additional cost, no data logging, no rate limits, no downtime, no censorship, no ads Think about it qwen 27b, 3090, Hermes agent, tailscale, curated textbooks

@tcarambat @PrismML Went back and ran llama.cpp and llama-bench using the prism fork. Nearly same speed to 1bit, just more memory:

Watched your video, thanks so much for sharing your thoughts! I will follow your demo to try the ternary model soon. I have used 1 bit inside LM Studio on 7800x3d, 32gb ddr5, 9700XT gpu. Got ~50tps. Prefill probably matters more with large contexts for agent harnesses with all the tool defs, etc. My prefill tok/s by measured prompt size: 1589 prompt tokens → 831 tok/s prefill 1757 prompt tokens → 838 tok/s prefill 3351 prompt tokens → 854 tok/s prefill 3606 prompt tokens → 831 tok/s prefill 6836 prompt tokens → 849 tok/s prefill 13335 prompt tokens → 803 tok/s prefill 26634 prompt tokens → 714 tok/s prefill 28680 prompt tokens → 713 tok/s prefill 56002 prompt tokens → 589 tok/s prefill Overall, seems very nice. I have vram and resources to spare to run TTS models on the side, web surf, and watch youtube! I like the direction they are going. Time will tell on capability.

You can now run Qwen 3.6 27B on <10GB of memory and retain 90% of the intelligence. Today, @PrismML dropped Bonsai 27B - which takes their unique compression approach they have already successfully applied to Qwen3.6 8B and Flux2-Klein and applied it to a properly smart model. Here is it using an entire computer for deep research. Long horizon agentic calls already are strong with Qwen3.6 27B and they still are here - just on 10x less memory.

@ingelramdecoucy It should be a dial, like volume control.

Perhaps unpopular opinion but I dislike the new algo, it’s not exactly that I don’t want to see my mutuals in my feed but Following already existed for that. For You’s usefulness was that stuff that was rising into the conversation tended to be pushed into it, which is broken now

@NoemiKhachian Why are badges getting dropped?

humbling experience not having my badge anymore but don’t worry i still work here

@MiaAI_lab I guess i have to buy a 50 series gpu now

NVFP4 keeps getting better and better.

@Dominos_UK No more pizza

tell a terrifying story in 3 words

@growing_daniel We need a dial

Lobster too buttery etc but literally every tweet on my for you now is a mutual, like what’s even happening on the larger website

the nature of careers have drastically changed my father worked in 1 industry; tech is my 4th he had 1 specialty his whole life; my relevant skillsets are constantly evolving what will it be like for our children? one thing is for sure we can’t even fathom it right now

@cremieuxrecueil I skimmed an article about this paper earlier today and I'm ashamed to say this didn't flag in my brain. I guess I just assumed people are getting dumber.

The study's sample also just seems, frankly, to be unbelievable. You're telling me you have an undergraduate sample that's less intelligent than typical adults?

This is a garbage study. They did the experiment a first time and the effect was obviously not real: 31.05 IQ points from just telling people a slight lie about how well they slept. In experiment 2, that dropped to 12 points. And then it failed to replicate.

@trademetry @OwenGregorian "every [good] thing in moderation" strikes again

@OwenGregorian 5 hours they are having fun...20 hours the kids are driving them crazy and stressing them too much.

Backwards science.

@michaelmalice how long does he have to make it to avoid a special election?

The reason we aren't getting specifics on Mitch McConnell's health condition is because his wife harvested his organs and you know it.

@DanielNealAdler This feels like as good a reason as any for me to start taking the smart watch off at night. Checking my "sleep score" every day probably isn't doing me any favors.

placebo effect remains unbeaten

@sheriyuo I liked this part the best: "reducing generation length by 19–46% with at most 0.7pp accuracy loss."

A 35B-parameter MoE agentic model with only 3B active that claims to match or surpass 100B-class models through post-training alone (no extra pretraining compute), via scalable agentic RL environments and a unified RL/OPD training stack with dynamic multi-teacher scheduling (17% training speedup). Mach-Mind-4-Flash Technical Report Paper: arxiv.org/abs/2607.09375

@TheQuartering @TheBabylonBee Big Cándido?

@TheBabylonBee By her own investigative standards she is also a man.

Candace Owens Confirmed To Be A Jewish Plot To Make Antisemites Look Retarded

@Timcast so does he still have to pay taxes on that, or?

Holy shit

@Endure2Thrive Depends on the assets and how much cash flow

What’s more valuable? 💵 Cash in the bank (1-2 million) or 🏘️ Assets that produce cash flow

@JustSayinFFS @AutismCapital Yeah, this one surprised me a bit.

@AutismCapital They did? I honestly thought these two were the type of people who could make it work

Everyone saw this coming.

@TheJoeySwoll Im also in my "something is better than nothing" era. Yesterday I got off my butt and did something not at all optimal, but something with low friction *for me*. Last night I got the best night's sleep I've had in 3 months. Don't let perfect be the enemy of good.

39 minutes is better than NO minutes at all. 👊🏼