uncanny

a second Fable-class model has hit huggingface

[ HAPPY TRAILS! ]

2 months later, I understand perhaps 70% of what 4.8 says

coming out of stealth with a 20B seed round led by a17 and pinetree to reinvent the AI employees landscape

brutalist architecture sticking out like a sore

aka 'traverse' is all you need hackage-content.haskell.org/package/base...

if this list of requirements is taken literally, a company can have no "moat" and can only become a utility

putting mathematicians out of a job with AI is a rookie game introducing CEOBench,

(4/6) ✨ 2) Keep improving our science — evaluation should mature into a measurement science, with validity, reliability, and reported error. And we need open weights, open source, and open checkpoints.

TBaaS tower of babel as a service

We've opensourced a simulation-based distributed property testing framework for #rustlang (based on proptest) Use case : finding minimal counterexamples to "does it converge within T rounds?" WIP: checking traces of black-box tools like Jepsen . Open to feedback! github.com/unfoldml/pro...

30 minutes and 1 context window later ..

tired of seeing all these "impressive results" from s-tier models. oh you got glm6.7 to one shot that refactor? fable did that job so good? wow. amazing. can you build for haiku or gemma tho

getting to the point, "theory of mind" of the interlocutor, etc are easy to spot, very hard to pin down could be that also explains the difference between benchmaxxing and optimizing for user comfort

reminds me I also have an experiment to share around JAX/OpenXLA .. sometime Soon

you see this with frontier models too, eg large parts of the response are "reheated" versions of the answer

there are serious risks involved with AI. for example, what if the most capable models fall into the wrong hands, who wish to use them to concentrate power, lock down access, hand unchecked control to authoritarian governments, and do secretive fingerprinting to sort users based on national origin?

this is pre-exponential thinking it will be ASI testing the gatekeepers for competence

anybody with a self-hosted GLM 5.2 who can run this prompt, would be great to know if it's a tuning vs prompting thing

would be fun to make one with a crank for power and a little steering wheel for alignment

Meanwhile memory chips are increasing in price so fast that by the time Neuromancer comes out this fall you will *actually* be able to get killed in Chiba City for 3 MB of hot RAM.

good thing everybody + dog has stocked up on mac minis for their janky calendar agent

Wiener would get a stroke if he knew an LLM is being prompted to act as a sensor

Next Monday, the EU decides on Chat Control 1.0. Will they let the Trojan Horse through the gate & let Google, Microsoft, LinkedIn etc continue to scan? We at Tuta say NO. ❌ No scanning ❌ No mass surveillance ❌ No age verification Call your local representatives! 🥊 #Privacy is not negotiable. ✊

love the setup with the mirror btw. Great for interactivity and selfies.

little ghost gets it

who's doing AI data curation and model evaluation on #atproto? show of hands and/or urls grazie cc @mariaa.bsky.social @bmann.ca

calling all #atmosphere AI builders, come one come all !

#1987 h1.nu/1xoY2

homemade Vietnamese spring rolls (my first attempt!) #foodsky

in research, even if you don't care about the lineage of ideas, you should have the maturity to discard 99.9% of what AI generates because it has been cribbed from prior art this matters less for code, but for scholarly work it's a disaster

first time I use conformal calibration in a real project. The quality of the predicted intervals is a bit hit or miss (depends a lot on the base model), but overall it's a great tool to master

The alpha version of my new book "Optimal Transport for Machine Learners" is out, with in particular an online version with interactive figures www.gpeyre.com/ot4ml/

gm

if Fable were AGI it would have found ways to circumvent the ban, just saying

using opus 4.8 to reimplement MiniMax M3 from scratch, come at me dario

same, but with a half-life of .. two days?

Not having any EU competitive labs (plural intended) is about to get very painful.

hold my kombucha

🥵🥵

another sober scenario for that "Europe 2031" essay that's making the rounds: China tunnels under Asia and Europe and steals ASML as a whole, factories and all

max plan: 100/month tens of extra CI hours due to an overzealous coding agent: priceless

Stop blaming yourself and others whenever something fails. Learn Feng Shui and blame the furniture.

lossy associative memory is both the best and worst feature of transformer LLMs

the old one-two: get users hooked, monetize

Court holds Google liable for false claims in AI Overviews. Seems significant. "A regular search engine just points to outside websites. But AI overviews generate 'independent, new, and substantive statements' by evaluating and combining content from various third-party sites" the-decoder

🥵