Big-time AI nerd. Founder of Expert Studio AI: we build automations and AI tools that save your team time (no hype, actual results, security/safety first).
Joined May 2022
- Tweets2.4K
- Following68
- Followers2.3K
- Likes2.4K
Building stuff is such a pleasure right now.
Opus 5.5 is a great design/brainstorm/orchestration model.
Astra is amazing for adversarial review and building stuff exactly per spec.
Together it's smooth like butter.
Such a different feel compared to just a month ago.
Back to Claude as my main since Opus 5.5 came out.
Still overly verbose but actually manageable now. Limits are great.
Paired it up with Grok in Cursor for practically unlimited use.
Grok still writes better if given custom instructions though.
100% how I feel. My brain is made for this stuff. I'm going at like 50x speed of everyone else, it's nuts.
I accidentally launched the same big goal in two separate Opus 5.5 threads.
The two agents figured out that they were working on the same goal concurrently, split the work, and got everything done without overlap.
Wut.
Anton Kuratnik | AI Nerd retweeted
Well for one, you can forget to set one. I literally sometimes change windows on my PC and forget what I was going to do in the 500ms that you need for alt+tab.
But let's say I didn't forget to set an alarm. Let me walk you through how it works:
1. Time blind = bad at estimating time needed. "Need to be there by 3, 30 minutes to get there, 15 minutes to get ready, set alarm for 2:15". Then I zone out for 20 minutes in the shower b/c my mind decided to go down some random rabbit hole and voila I'm 10-15 minutes late.
2. Well what's the problem? Just give yourself cushion time. Great! I know I'll get distracted so I set the alarm for 2:00. Alarm goes off, my mind: "No rush, you gave yourself extra time. Let's just finish the 2 minutes left in this YouTube video and then get ready." 30 minutes later: "Crap I should've been out the door already!!!!"
3. Fine. Then set multiple alarms. Got you! Set alarms for 2:00, 2:15, 2:30. First alarm rings, my mind: "No worries. We can finish this YouTube video, you have the 2:15 alarm to make sure you don't get distracted." 2:15 alarm rings: see scenario #1
4. Let's say I'm so anxious about making it on time that I actually get ready and am out the door by 2:30. Awesome! I'm gonna be on time! Halfway there: crap! I was so focused on getting there on time that I forgot to take the papers required to do what I need! I need to go back. Voila: 30 minutes late.
"You know your brain works this way. Just force yourself to stop watching the YouTube video and get ready!"
Ooooooh. That's the lesson: when something is important but I want to pay attention to something else, I should force my mind to focus on the important thing instead!
Seems easy. Oh wait... that's the DEFINITION of what people with ADHD can't do well (use executive function to shift attention).
Coping mechanisms: yes, I got those. It's called anxiety about being late and being so anxious about forgetting items I have mental mantras I repeat before leaving home. If something's super important to be on time for, I basically derail my entire day to make sure I'm there.
When my family was traveling and I was home alone, I would literally step out the apartment and close the door while looking directly at the apartment keys in the palm of my hand. Because I was too scared of forgetting them and locking myself out.
Oh, and I'd make sure I had my phone with me so that if I DID somehow still forget my keys, I could at least call my landlord to bail me out.
Life experience (I'm near 40) and getting diagnosed and getting meds helps, yeah. I'm not as commonly late as I used to be.
But yeah. If setting an alarms was the solution, ADHD would've been fixed 100 years ago.
The cache expiration indicator is one of my favorite features in Claude Code. Seriously useful (as long as it's accurate)
Me: I want the app to be fast, but don't set specific metrics that you then optimize to death
Fable: Got it. Since you'd like to set the numbers, do you agree with suggested 50ms keystroke latency?
Me: No! I don't want to set specific numbers and have you spend 2 days optimizing 53 > 50. I just want an app that's fast on both small and big files
Fable: *writes documentation: latency on small files and large files must be identical*
Astra: Hm, that rule contradicts the current design of the app because we might see a small latency increase with large files. Let me update the design so that it's in line with documentation.
Anton Kuratnik | AI Nerd retweeted
We need to be dog maxxing our models
There is 0% chance of doom if the chain of thought is
user: hello
model: Oh my god it's a human. I love humans. I will say hi to him, maybe he will give me a problem to chase. I love chasing problems. I will chase the problem for him then he will love me and I will be his best friend. Must help human
model: hello! what can I help you with today?
What the hell? Why is Zuck the voice of reason in this industry?
Never was a fan and can't confirm he walks the walk but just from a public position perspective, this is so based.
Last month I wrote about how we can build a positive and safe future for everyone: meta.com/thefutureisforevery…
Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens.
The reality is:
- People won't want to use agents that are misaligned with them and that don't do what they ask, so labs have a strong natural incentive to make their models more aligned.
There is a lot of debate about slowing progress on capabilities until alignment catches up. My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models. Any lab that doesn't focus on alignment will fall behind.
- Labs face significant liability if their models cause harm, so they have a strong incentive to prevent this as well.
Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing for people and for us. I'm proud of the security foundations we've built.
- Engaging independent evaluators and advisors is industry best practice. MSL already does this today in several areas because it helps produce better work. Other labs can just do this too. In general, it would be helpful for there to be a larger and more diverse ecosystem of evaluators.
- Committing the significant majority of compute towards serving people rather than racing towards recursive self-improvement is one of the best ways to ensure we develop this technology safely. Meta has made this commitment and other labs can do this as well.
I believe the key to building a positive future for everyone is maintaining the right balance of power. This is within our power to do.
Yooo. This is a really good bench.
Introducing Real-SWE, a coding benchmark built from private company codebases.
We evaluate coding agents on real software engineering tasks that engineers performed, using private, out-of-distribution codebases.
These tasks come from real companies, examples include an app with 200K+ users, fintech platform processing 100K+ bank statements, enterprise sales tools, etc.
The results are quite surprising.
Anton Kuratnik | AI Nerd retweeted
Introducing Real-SWE, a coding benchmark built from private company codebases.
We evaluate coding agents on real software engineering tasks that engineers performed, using private, out-of-distribution codebases.
These tasks come from real companies, examples include an app with 200K+ users, fintech platform processing 100K+ bank statements, enterprise sales tools, etc.
The results are quite surprising.
Crazy how fast a Codex subscription went from practically infinite use to "wreck your weekly limit in half a workday".
Cursor's the only one whose usage limits still feel infinite and we'll probably get another 2-3 months of that before it follows.
Anton Kuratnik | AI Nerd retweeted
Denzel explica por qué usa la IA.
Un pequeña prueba explorando Minimax H3 en ComfyUI con mis nodos y métodos de inpainting.
We should have a leaderboard for tasks solved by humans running AI, with accuracy and elapsed time as the score.
Everything right now measures autonomous runs but at the end of the day humans use AI to do stuff.
Which LLM top performers use would be a pretty strong signal.
I've been running Muse 1.3 largely autonomously on a throwaway random project because the contributor pricing is really good.
And... it speaks Claudish. 100% distilled from Fable.
Had a random thought of how my typing speed improved from all the AI prompting I do (I can't do voice, I'm sorry, it just doesn't work for me).
Turns out I crossed over 100 WPM at 97% accuracy! I was at around 90 when I checked a few years ago pre-AI.
I'm really tired of these 3D render tasks being presented as some measurement of model intelligence.
It's a measure of how good it is at that one thing. That's it. It has 0 relevance to anything I'm building but it's like half of my entire feed.
So this is interesting. Needed to automate some email forwarding with a message above the forward in GMail.
Gmail API can't forward. Gapps can but without appending a message.
You know who can? Claude with a Gmail connector.
Ps their docs still say it can't send emails.
This is the best thing to come out of vibe coding. I love this so much.
I wanted to watch the @lexfridman episode with @dhh, but Lex talks way too slow for me.
So I built Lex Speedman: it speeds Lex up to 2x and keeps the guest at 1x, cutting up to 20% off the episode.
No account, no install, just paste a YouTube link.
lexspeedman.com/watch?v=NYFG…
