Meet Husky: a Model-Specific Inference (MSI) engine up to 4.5× faster than Apple's MLX
Woof, Underdog's Pareto frontier model, now runs up to 730 tokens/sec on a MacBook
Finally local models are as fast & capable. Try it now in underdog.ai - your personal private AI
259
221
2,749
427,480
I've used Underdog for a few days, and it's the best consumer AI experience I've had since ChatGPT.
It's a genuinely useful AI that looks at your email, calendar, meeting notes, etc. But, most importantly, it's 100% local. You own your data!
Why I love it ⬇️
3
7
25
23,687
@0xSigil is on to something great. Also incredibly responsive, I've filed ~20 tickets since Friday and the majority have already gotten shipped to prod.
There are still small things that can be improved, but the app is getting noticeably better every single day.
1
4
260
This isn't a paid post also, I just love the app. I consider myself an early adopter in tech, and this is genuinely the best experience I've ever had with an app so early.
Seriously, go try it out @underdogdotai ! You won't regret it
2
1
9
326
Beautiful tweet, has led me to discover some wonderful personal sites. Here's a list of my favorites:
nat.org by @natfriedman
jrands.com by @jayrizpop
benji.org by @benjitaylor
ped.ro by @peduarte
commissioner.design by @ukauwa_david
sasi.codes by @sasicodes
murphyusifo.com by @druids01
alextudun.com by @AlexTudun
sanyi.space by @imsanyi
timourkosters.com by @timourxyz
17
32
780
132,096
This cost OpenAI ~$25m in compute. Here are 4 alternatives that would be 10x better in improving their public image:
1. Diagnose rare diseases. Millions of Americans have rare genetic diseases, and many never get a diagnosis. However, there are hundreds of new gene-disease pairings published every year. Studies have shown that ~15% of unsolved cases get solved upon re-review with the latest research. Unfortunately, it's incredible expensive for a human to do this. But, OpenAI could do it cheaply with agents, and pass off a final report to a physician. This would instantly improve thousands of lives.
2. Get veterans their benefits. There are hundreds of thousands of veterans with unprocessed benefits. It's very challenging to put together the relevant information and file properly. There are volunteer organizations that help with this, but it's not enough. In fact, there is a shadowy industry of "claim sharks" that will help veterans file their benefits for a cut. OpenAI's agents can easily help every veteran file their claims in exactly the right format.
3. Translate the world's textbooks. There exists a large corpus of free, excellent educational material (MIT Open Courseware, Khan Academy, OpenStax), but translation to other languages has been flaky. OpenAI could very cheaply translate ~all necessary educational material to every major language in the world, and pay for great human review.
4. Fix government forms. Americans hate filling out government forms, and their complexity leads to millions of error annually. OpenAI could easily convert every form to a human-legible version, and pay for lawyers to review the final output. It's possible they could even partner with the government to make this an endorsed channel. Would improve their public image with literally everyone.
Frontier labs have access to unlimited capital and superhuman intelligence. But they have the worst public perception among all major companies. They should apply their capital and intelligence to fixing that!
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
3
9
1,199
To be clear, it's incredible that frontier models are able to solve deep, open problems in Mathematics. And they should keep doing so!
I just wish that the Labs would spend a similar amount of effort in helping improve humanity. For companies creating such an incredible amount of technological progress, they've done an awful job at convincing normal people that it will be good for society
1
136
Why don't speed cameras ticket intelligently?
It's bad that humans and cameras act differently. Cameras should mimic the best humans, and consider traffic, road conditions, etc.
Automation should scale what works, not introduce a new policy that the law was not designed for.
I usually dislike speed cameras. I accept that there should be a law but think that appropriate speeds are determined much more by road conditions than by posted signs, and notice that almost everyone speeds a lot in many places where posted speeds are too low.
I do not trust the government to set speeds efficiently, and I resent being punished for maintaining the flow of traffic, particularly in areas where I’m passing through for the first time and get surprised by unknown traffic cameras. I do not think most speeding should be ticketed.
I strongly prefer speed laws be enforced by humans with discretion and judgment who can tell the difference between someone recklessly weaving between every narrow gap and someone going 15 over on a straight, empty stretch of highway with no real danger to others present.
1
463
My takeaway is that Instinct is going to sell user data
Instinct needs a $100B+ outcome to justify a $2.5B val
But, consumers don't pay for AI. ChatGPT has 1B users, and only makes ~$8b/yr from user subs
The only path to enough revenue is monetizing user data
Instinct, the viral AI assistant, has already raised $350 million, including a new round that values the startup at $2.5 billion. wsj.com/tech/ai/the-latest-v…
57
6
315
57,347
The market for data is much bigger than the market of paying consumers, so I expect "data selling" AI assistants will become much better funded (and have a better product) than "private" AI assistants
2
9
2,457
I've been seeing a lot of posts lately advising folks to not switch models mid-session.
Despite its ubiquity, I don't think this is advice is correct.
Switching models costs, at most, $10. Often you will save > 30m of your time by switching to a stronger model. A good trade!
Just FYI. If you are switching models in a session all the time - you are doing it wrong.
Every time you switch, your entire prompt cache is invalidated on the new model you switch to, and you have to repay the full input tokens price for all of it.
Stop doing this unless those models are free.
This is not a hermes thing - this is a fundamentals of inference thing.
2
2
951
Thoughts from first day using Instinct:
It's really nice that I got an invite! I love being part of this exclusive group. If you want an invite, don't bother. If you were cool you would have already gotten an invite by this morning (when I got one)
6
1
30
4,084