let's go exploring
New York, USA
Joined January 2017
- Tweets489
- Following519
- Followers2K
- Likes4.6K
Laws that would make NYC more livable:
- $1 tax per honk
- Ban tourist helicopters
- License plates for e-bikes
- Increase congestion pricing
- Make sirens quieter
- Allow Waymos
- $1,000 fine for sidewalk driving
- Escalating fines for parking violations
Another, more powerful, example: Astra can drastically improve the performance of tiny Qwen by hand-editing its weights.
This is after a team of highly specialized researchers already worked on optimizing Qwen
5/6 Can an LLM improve another LLM?
We simply asked Astra to make Qwen3.5-0.8B better at math.
Astra saw Qwen get stuck in reasoning loops, which is a known issue that Qwen’s own model card warns about (see their official model card). With a budget of 50 forward passes on practice problems, it located the cause and edited it out. This used no gradients or training.
Fixing these loops by editing just 0.016% of weights boosted other capabilities beyond math too!
This whole thread is magical. Models are now better than humans at understanding how models work. Very meta!
A delightful example:
If you want to make a NN that can play BreakOut, you're better off asking Astra to hand-create weights than RLing your own model for ~1 GPU-hr
Hot take: this is good. It's bad for humans to be cruel and vindictive, even to inanimate objects. It reinforces antisocial behavior.
Users are also clearly doing this because AIs feel human-like. No one engages in sustained cruelty against their toaster.
Exclusive: Anthropic is updating its usage policy for the first time in over a year. The new rules prohibit sustained "abusive or cruel behavior" towards Claude & add new restrictions about propaganda campaigns, surveillance & weapon development. theverge.com/ai-artificial-i…
I've been reflecting on how AI has transformed our lives over the past few years.
Subjectively, I think there have been 9 step-changes in how AI is used:
1. Nov 2023: ChatGPT launch
2. May 2024: Google AI Overview. This made AI accessible to everyone, and remains the most common way users interact with AI.
3. Oct 2024: GPT-4o launched web search. AI becomes the front door to the internet.
4. Apr 2025: OAI o3 launch. This was the first good reasoning model. Users were able to replace tasks that took 3-5 tabs with just one query.
5. Nov 2025: Opus 4.5 launch. This is the point where coding agents became viable. When "vibecoding" really took off.
6. Jun 2026: Fable 5 launch. This is the first time models had reasonable judgement. AI can do a good first pass at almost any task.
7. Jul 2026: Grok 4.5 launch. This is the first capable model that was extremely cheap. Small tasks become effectively free.
8. Aug 2026: Instinct launch. This was a substantial UX and harness improvement, and made agents accessible to the average person.
9. Sep 2026: Opus 5.5 launch. The first model that combines judgement, skill, and speed at extremely low cost. Most coding tasks become ~free.
Two observations:
(a) advancements on the frontier have not led to persistent advantages. Any new model capability is generally available within 1-3 months.
(b) the rate of change is undoubtedly increasing. Every few months now, I fundamentally change the kind of tasks I can do with AI.
Are there other moments that changed how you use AI?
Hark is incredibly polished, kudos to the team for building a great product
But, imo the launch needed a one-line differentiator
Muse: cute (best hook by far)
Instinct: lives in iMessage
Grok Bot: AI teammates
Dots: OAI-native
What's the one-liner for Hark?
Introducing Hark Pro
Hark Pro is now live on web, iOS, and Android
Try it free: hark.com
Underdog is the best consumer AI experience I've had since ChatGPT
Introducing Underdog, your Private Personal AI on devices you already own
Today we're announcing our backing from @a16z @khoslaventures @HummingbirdVC Anthology (@AnthropicAI @MenloVentures) @patrickc @naval @rauchg @polynoamial @Thom_Wolf @Mascobot @tszzl @OfficialLoganK and other top AI leaders
Our mission is to provide free, capable, reliable and private AI to billions of people.
Underdog’s Law: today’s frontier intelligence reaches your devices in six months.
We’re starting with fast, capable AI that runs on consumer hardware. By co-designing models and inference engines, we’re pushing the frontier of capability, speed, power and data efficiency.
Our research also spans agentic commerce, confidential inference, and how AI will reshape the internet economy.
Privacy and capability no longer needs to be a tradeoff. If you believe in this future, join us. Time to build.
Vishal Talasani retweeted
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
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 @underdogai ! You won't regret it
@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.
Importantly, the model is incredible. I'm not sure what kind of magic was used to train it, but the AI takes up ~2.5gb of RAM, and runs at ~40 tok/s on my Macbook Air.
In terms of quality, for the vast majority of my queries, it's just as good as Claude or ChatGPT.
I've already deleted Superwhisper from my Mac, and I'm probably going to switch from Superhuman to Underdog soon.
I've never used an AI notetaker because of the privacy concerns, but I've also started using Underdog for notes.
It's also now my default Calendar app.
Underdog changes all of that! It's a true superapp, it currently combines:
1. Chat (Instinct)
2. Email (Superhuman)
3. Meeting Notes (Granola)
4. Transcription (WisprFlow)
And more, all in one app. The UX is terrific, a nontechincal user could easily onboard in < 3 minutes.
For a long time, I assumed local models weren't good enough to do this. I've tried a few local AI tools, and they're mostly terrible. Both (a) the models aren't quite powerful enough to do what you want, and (b) the UX to access those models is below par.