Ehsan Azish
1.4K posts
Ehsan Azish
@ehzish
Indie app maker, Machine Learning specialist, iOS developer, swiftUI
Katılım Aralık 2023
650 Takip Edilen283 Takipçiler
@TurnerNovak half a seniority level flatter is just a fancy way of saying nobody junior got hired
English
Interesting research on how an "AI-native company" looks different
Using data from YC companies as a sample, AI startups are smaller, flatter, and more expert dense. They have a greater shares of engineers, more folks in senior roles, and fewer managers.
Startups that use AI for processes look the same. But startups that embed AI directly into their products are organized differently.
Biggest impact is found in service-based firms, especially healthcare and education. These businesses historically scaled by hiring more people, which you can do more efficiently with an AI product.
The data is from early 2025, so doesn't include any impact from Claude Code or Codex.
Rem Koning@orgRem
🤖Excited to share a new working paper.🤖 The phrase "AI-Native firms" it everywhere, but are they any different? Is it just hype? In our new paper we show AI ventures are organized differently, but not for the reasons you think.
English
@theo the changelog cuts off right as it gets interesting, that's the real transparency test
English
Unprecedented levels of transparency out of OpenAI, really cool to see. They definitely made mistakes but they’re addressing them with full transparency
Tibo@thsottiaux
Updates for Codex and ChatGPT Work users. No nerfing, only good stuff! - We have landed inference optimizations and are passing down savings to all the subscriptions for GPT-5.6 Sol. That should result in around 10% more usage on its own. - We noticed that by changing the context size limit in the product to 372k for GPT-5.6 Sol, up from 272k for GPT-5.5, it resulted in more usage being charged than intended. We have reverted to 272k and will work to roll back out to 372k in the days to come. You should notice that usage drains significantly less after this change. - To understand where the extra usage was coming from, we ran some experiments where reasoning efforts were changed (referred to as juice values under the hood) and have reverted this. - There is slightly more usage of multi-agent than intended in high and xhigh reasoning effort, we are fixing this going forward. Also fixing a small other thing we noticed with auto-review where we can be more efficient. And we continue to have the 5h limit temporarily not apply. Enjoy the rest of the weekend!
English
@burkov so the discriminator's job is basically being a professional nitpicker, my crash reporter wishes it had that job
English
In this CoNLL 2026 paper, the authors from from MIT develop an adversarial generator-discriminator framework that effectively integrates human demonstrations with verifiable rewards, enabling large language models to optimize for both objective task accuracy and subjective, non-verifiable human-like qualities: chapterpal.com/s/da21ca57/rig…
English
@dani_avila7 not solving it, just avoiding admin policy entirely and shipping single binaries per client
English
How are you solving this? Seeing the same problem in Codex
Any tips to improve plugin distribution across clients?
Daniel San@dani_avila7
Why don't marketplace plugin selection policies mapped to SCIM groups in Claude Admin work in Claude CLI? This is a critical enterprise requirement, It breaks the entire context administration model for org-wide Claude deployments The flow is well designed, but this one gap breaks the whole rollout CC @bcherny , @trq212 , @amorriscode
English
@Austen meanwhile some of us are still manually writing comments nobody reads
English
Was laughing at dinner tonight about how ~18 months ago someone asked me if they thought you could use AI to write comments for all of their code; they didn’t trust it to write the code yet.
Turns out YES!
OH BOY CAN IT WRITE COMMENTS!!!
English
@thsottiaux juice value 960 down to 128 and they're calling that a wash
English
Updates for Codex and ChatGPT Work users. No nerfing, only good stuff!
- We have landed inference optimizations and are passing down savings to all the subscriptions for GPT-5.6 Sol. That should result in around 10% more usage on its own.
- We noticed that by changing the context size limit in the product to 372k for GPT-5.6 Sol, up from 272k for GPT-5.5, it resulted in more usage being charged than intended. We have reverted to 272k and will work to roll back out to 372k in the days to come. You should notice that usage drains significantly less after this change.
- To understand where the extra usage was coming from, we ran some experiments where reasoning efforts were changed (referred to as juice values under the hood) and have reverted this.
- There is slightly more usage of multi-agent than intended in high and xhigh reasoning effort, we are fixing this going forward. Also fixing a small other thing we noticed with auto-review where we can be more efficient.
And we continue to have the 5h limit temporarily not apply. Enjoy the rest of the weekend!
Fix@FixlationAI
OpenAI has reduced GPT-5.6 Sol's thinking budgets in an effort to make the model more efficient They essentially bumped everyone's reasoning down by 1... so if you were running Sol Extra High, you now have to set it to Max to get the same effort So we basically don't have Max reasoning anymore, how do you feel about these changes? 🤔
English
@MatthewBerman burma superstar still holding it down or has that changed too
English
@natolambert anthropic just quietly good at the boring knowledge work tasks nobody hypes
English
Claude Fable is another big step in being able to make nice lectures based on existing educational content. Much better than Opus.
GPT 5.6 is still very far off here. Is a good example of where Claude Code being a bit easier to work across different knowledge work tasks.
English
Most people think Harry Potter is the best selling book series in history at ~615M books.
That will likely change this year.
It will be overtaken by One Piece, by Eiichiro Oda, regarded as the greatest Japanese manga / graphic novel of all time, which just crossed 600M in March and is projected to surpass HP this year and hit 725M+ when it ends in ~2030-32 (there’s 10-15% left).
The story follows Monkey D Luffy, a young boy who has rubber man powers, who assembles a pirate crew and goes on an epic journey to find the legendary “One Piece” treasure.
As a voracious reader, I struggled to take manga seriously for most of my life. I’m a hardcore Harry Potter fan, and even thinking of manga in the same breath once disgusted me. I relegated manga to a “niche comic book for nerds” and anime to a “kids cartoon”. “These aren’t real books, it’s just for people who can’t read longform”, I thought. But the past few years have completely changed my mind. I’ve been personally reading and watching (YouTube/Crunchyroll) many manga / anime including One Piece. Aside from the appreciation for the artistry of the characters, the stories strike me as immensely powerful with deep lessons about loyalty, sacrifice, friendship, ambition, justice, power, grief and kindness. Once you get in the flow, it’s far more enjoyable than most non-fiction and rivals anything on my shelf. The storytelling is beautiful. It awakens the child inside you without being juvenile. I highly recommend it to anyone who has previously had reservations like I once did, and doubly so if you want to enjoy something with your kids.
English
@vkhosla sounds simple until someone adds a parking minimum back in
English
How about a California initiative that allows for minimal restrictions and no zoning restrictions other than meeting the basic state building code on any housing that’s targeted at the lowest quartile of housing prices in the county. Works for every county in the state. Short time, construction and lower cost at the low end. Will increase supply at the low end of prices.
Garry Tan@garrytan
They’re going to blame tech But the real cause: extreme restriction of housing supply and subsidized demand
English
@dani_avila7 that red arrow is doing more honest reporting than most changelogs
English
Why don't marketplace plugin selection policies mapped to SCIM groups in Claude Admin work in Claude CLI?
This is a critical enterprise requirement, It breaks the entire context administration model for org-wide Claude deployments
The flow is well designed, but this one gap breaks the whole rollout
CC @bcherny , @trq212 , @amorriscode
English
ChatGPT 5.6 Sol (Pro) explains the MIT License in one paragraph using only words that appear in public domain songs sung by 1920s British ballroom sensation Al Bowlly
English
@XFreeze reconstructs a whole street out of a glare blob, meanwhile my app still can't reconstruct why it crashed
English
Tesla Vision is one of the most advanced perception systems in the world
FSD processes raw camera pixels through an end-to-end neural network and turns them directly into driving decisions
In difficult conditions, its photon-count reconstruction allows the car to see through extreme glare and darkness......even revealing clear details that are nearly invisible to human eyes
Just cameras, massive real-world training data and rapidly improving AI
Tesla is teaching cars to see the world at a level humans simply can’t match in certain conditions
English
@DeryaTR_ exponential for the models, linear for the guy still profiling his own app at midnight
English
Can you imagine that just 2 years ago today the top AI model was GPT-4o & we had none that could do coding! In a year or so from now, when we look back to today, we will say similar things! This is why it is almost impossible for most to imagine the exponential advance of AI age!
Serafim Batzoglou@s_batzoglou
The status of the AI competition is incredibly fun to watch and the most exciting technological race since... the dawn of humanity: Fable 5 was a big event. Superb intelligence. But OpenAI played it much better imo. Anthropic is melting a lot of trust and good will with all the drama. OpenAI released Sol (a Fable competitor at lower price), Terra (slightly behind Fable), and Luna (better than GPT 5.5 in reasoning!) and quickly responded to complaints of quickly vanishing usage limits. OpenAI is very good with timing. Back last fall, they released GPT 5.1, got Gemini to "unload its guns" with releasing Gemini 3 pro, and then quickly released GPT 5.2 to beat Gemini. The most interesting development, however, is the broadening of the race. Open models DeepSeek 4, Kimi k2.7 and GLM 2 (the latter of which I don't like much) brought lower prices and, well, open models! Good for fine tuning and mechanistic interpretation. Big pressure on price. And now we have two new entrants close to the frontline. Grok 4.5, finally a big improvement on previous Groks: a small improvement over Grok 4 with better price and more reliability, a huge improvement over the weak and nonsensical Grok 4.3. And most impressively, Meta Muse Spark 1.1. A real contestant, bringing GPT-5.4 level reasoning at a very low price, and making us anticipate frontier performance by Meta. A very very exciting time and race!
English
@natolambert nsf spends years building networks then torches them with a memo
English
Another insane own goal on AI research leadership from the U.S. It's like "how do I remove myself from an international scientific network" 101 class
Kyle Chan@kyleichan
Utterly self-defeating. “The National Science Foundation (NSF) has decided to ban collaborations between every U.S. scientist it funds and nearly all Chinese research institutions and their employees.” science.org/content/articl…
English
@fchollet my crash logs are basically a diary of failing gracefully, adapting less gracefully
English
The process of trying, failing, updating a mental model, and trying again is the core of intelligence. We should celebrate models that fail gracefully and adapt instantly.
English
@TrungTPhan first take vocals on a song that big, meanwhile i'm still rewriting a readme for the third time
English
Oasis just beasts in mid-90s.
This clip of them making “Champagne Supernova” in 1995 is gold.
Noel Gallagher writes the lyrics. Does a single play-through for his brother Liam. Then Noel goes to watch football while Liam — after one listen — smashes the song on first recording.
Trung Phan@TrungTPhan
Only two songs from the 1990s have over 3 billion streams on Spotify: Goo Goo Dolls "Iris" and Radiohead "Creep". Oasis "Wonderwall" on the cusp and almost certainly joining that list after the World Cup.
English