Pinned Tweet
I'm excited to announce that The White House will be doing a Global Accessibility Awareness Day (#GAAD) Event and Town Hall tomorrow at 10AM EST.
Links to Livestream and Access Streamtext in the comments.
Thanks @WhiteHouseOPE46!
From blog post to The White House. 🤯
Links to Livestream and Access Streamtext in the comments.
Thanks @WhiteHouseOPE46!
From blog post to The White House. 🤯
12
19
81
13,951
People suck.
Vulnerable post: I spend 90% of my time on @chatprd dealing with increasingly creative international token fraud operations, and it makes me want to shut down the company. Mostly to steal Opus tokens on resold promo codes or in the time between a trial start + our fraud alerts infra.
It's cost me tens of thousands of dollars and it gets worse every day. Luckily the company is nice and profitable, but every time it happens it bums me out in extreme ways.
I love ChatPRD, I think at this point we've helped PM, engineers, and founders with millions of products + features. We are still a top 3 writing GPT (completely free!) on ChatGPT, we've kept generous trial periods, and have kept annual sub costs super reasonable, even for the enterprise. We've let people use the latest models when they come out, and have built in new features like prototyping without adding to our fees. We haven't had to limit usage, because our customers have been so great.
Building the product is incredibly fun, it's an actual footprint by which I get to test all this AI stuff, as an engineer, founder, and product builder. I have my little grok bot army that runs ops. We've been pen tested multiple times, SOC2ed, and work with some of the biggest companies out there.
It's my labor of love, and I keep it going (even though there are probably higher leverage things I could be doing with my time.)
But it's horrible to have our small team spend basically all of our time shutting down that Indonesian fraud ring, or that Egyptian code farm, or the Chinese signups that have been on an auto loop for the last 24 hours. I feel like I'm constantly making product calls that trade protecting from abuse vs giving our users an amazing experience (and keeping maintenance overhead low.)
It's basically sucked the joy out of my first little AI win, and I'm really sad about it.
And before you bucket me into dumb-little-vibecoder, trust that I know what I'm doing. I run our firewall like it's my full time job. We are stripe radar'd up to the gills. We have rate limits + so many automated triggers and controls.
But each one of these things ends up touching legitimate users and prospects. And every time I rachet up one control, I end up causing annoyance or pain to a real user. It's a trade that's hard to manage.
And honestly? It's not fun.
It just feels bad to see so many bad people.
So I have some thinking to do, but for now
- free trials are dead on chatprd, sorry kids
- i think in 2027 i might ban all non-US business
- and tbh, may retreat to the warm waters of 100% enterprise business, who are a) willing to pay more and b) have more interesting product needs
And if you're a little solo founder or startup that has dealt with this - I see you. I have some tips + tricks that might help. You're not dumb, and it's not easy.
Ok. Rant over.
Back to building the little wrapper who could 💕
0
0
1
42
Looks like Kevin extended @dhh's Once (campfire) slack clone. This is very exciting. I got scared to deploy Once because @omarchy is using discord, but this gives me confidence. Thanks @kevinrose!
Want your own Slack for $0? I open-sourced Ember. It deploys to @Cloudflare’s free tier in about a minute, works with Slack-compatible payloads, and includes an AI assistant and media attachments.
0
0
0
104
This thread threw me for a loop. I don't understand something.
Before getting into it, I'll reference another thread you don't want to miss, before tying them both together.
There are two kinds of battles between techies. I hate seeing fights where people I respect attack each other personally and with vulgarity. I really hate when they pretend their enemy is unaccomplished.
The second kind can get a little heated, and even a few names thrown in, but fundamentally, it's a difference of philosophy. There's something to learn.
This thread is the latter. It's like the @RichardSSutton debate yesterday, where Richard, who coined the phrase the "Bitter Lesson", later says stuff that seems to contradict it, which upsets the folks who are Bitter-Lesson-Pilled.
I want to understand if Richard changed his mind, or feels that his essay was misunderstood. I suspect the latter.
Similarly, this thread features @badlogicgames, who literally wrote @pidotdev, one of the best and most important, open source, AI coding harnesses in the world. He is siding with a project that bans AI coding. You would think Mario would be AI-Coding-Pilled. But perhaps he is not.
He seems very skeptical of the quality of AI generated code. And I do get that. Personally I'm torn. I am all in on AI coding, but even today's models often write code that is dog-poo. You need to be technical and know how to steer these suckers.
But I suspect it will get way better. And unlike many who disagree with Sutton, I think we need a different paradigm before we get there. The @ylecun idea that we need World Models that "understand" the words rather than just predict the next word.
(I can hear the opponents arguing that LLMs do understand, and in a way they do, but Lecun and Sutton define "meaning" differently I think).
Mario, do you feel that this project is different? That it requires so much craft that AI is not appropriate?
If not, why write an AI coding harness if you think projects should ban AI-written code? I'm so confused.
P.S. I'm not tagging the target of this post, who created Omarchy, because Mario didn't. But it's inevitable he will see it and weigh in.
Before getting into it, I'll reference another thread you don't want to miss, before tying them both together.
There are two kinds of battles between techies. I hate seeing fights where people I respect attack each other personally and with vulgarity. I really hate when they pretend their enemy is unaccomplished.
The second kind can get a little heated, and even a few names thrown in, but fundamentally, it's a difference of philosophy. There's something to learn.
This thread is the latter. It's like the @RichardSSutton debate yesterday, where Richard, who coined the phrase the "Bitter Lesson", later says stuff that seems to contradict it, which upsets the folks who are Bitter-Lesson-Pilled.
I want to understand if Richard changed his mind, or feels that his essay was misunderstood. I suspect the latter.
Similarly, this thread features @badlogicgames, who literally wrote @pidotdev, one of the best and most important, open source, AI coding harnesses in the world. He is siding with a project that bans AI coding. You would think Mario would be AI-Coding-Pilled. But perhaps he is not.
He seems very skeptical of the quality of AI generated code. And I do get that. Personally I'm torn. I am all in on AI coding, but even today's models often write code that is dog-poo. You need to be technical and know how to steer these suckers.
But I suspect it will get way better. And unlike many who disagree with Sutton, I think we need a different paradigm before we get there. The @ylecun idea that we need World Models that "understand" the words rather than just predict the next word.
(I can hear the opponents arguing that LLMs do understand, and in a way they do, but Lecun and Sutton define "meaning" differently I think).
Mario, do you feel that this project is different? That it requires so much craft that AI is not appropriate?
If not, why write an AI coding harness if you think projects should ban AI-written code? I'm so confused.
P.S. I'm not tagging the target of this post, who created Omarchy, because Mario didn't. But it's inevitable he will see it and weigh in.
man, who should i trust? a bunch of nerds who made linux run on apple silicon, reverse engineering that pesky gpu isa among a gazillion other things, all pre-ai.
or the guy who thinks having your agent rewrite shitty rust he never looked at in x86_64 is excellence?
0
0
2
292
Dave Winer of RSS fame built a markdown based web browser. It is interesting. Good for accessibility. Many benefits. Doesn’t feel like it’s intended as a replacement for the WWW. Perhaps for blogs only?
Need to reflect on it.
Also will folks pick it up?
Need to reflect on it.
Also will folks pick it up?
Here's why I think the web of Markdown will not be taken over by silos. Short version, the same reason podcasting has never been absorbed that way. We are setting user expectations right from the beginning.
1. Users control appearance.
2. Writers control the words and punctuation. And of course links.
I wrote a piece summarizing what we accomplished yesterday, please read it if you're following this.
demo.bonbonweb.org/?url=https%3A%…
And I have another big thing to demo today. If you liked what we did yesterday, you're going to like this too. ;-)
0
0
0
154
Joe Devon retweeted
A cancer drug designer just told me AI makes his work ~100x more efficient and ~100x more effective.
Not some AI hype guy. Matt Soellner is a University of Michigan chemistry professor who's spent almost 20 years designing the kind of drugs that keep people like me alive.
A se AI models were silly. Now he calls the last 3 months the highlight of his career.
When a guy like that says 100x, buckle up. Cures are about to start coming at a pace nobody's ready for!
Not some AI hype guy. Matt Soellner is a University of Michigan chemistry professor who's spent almost 20 years designing the kind of drugs that keep people like me alive.
A se AI models were silly. Now he calls the last 3 months the highlight of his career.
When a guy like that says 100x, buckle up. Cures are about to start coming at a pace nobody's ready for!
Yesterday I posted about working with an AI scientist to find new ways to fight my cancer.
A chemistry professor at the University of Michigan saw it and emailed me. Matt Soellner @mbsoellner works on the exact kind of lung cancer I have. He offered to synthesize compounds and test them in cells and animals.
A lightbulb 💡 goes off above my head. Obviously...I MUST introduce Matt to Jan.
Andddd by this morning, they're talking about how to take an AI-designed molecule off a computer screen, make it in a lab, and find out whether it can fight my cancer.
EXCUSE ME!? Less than 24 hours after my post.
We haven't made a drug. We don't know whether any of this will work. Making and testing compounds will take months.
But Matt saw a patient trying to find a way forward and chose to help. He brought the ability to make and test things that might otherwise never leave a computer screen.
I'm sharing this as it happens because this is what gets me excited about AI. I can put my story out there, connect people with different skills, and help move an idea toward an actual experiment against cancer.
I deal with the every day bullshit from cancer as a patient but have realized there's an army of incredible people out there working hard behind the scenes on treatments and cures for people like me and millions of others.
It's an honor and a privilege to bring their efforts out for the world to see. AI is accelerating what these scientists can do. I'm just the human meat proxy making introductions and sharing the story.
But if that helps get a promising idea off a screen and into a lab faster, I'll take it! 🥩
9
12
153
4,346
Joe Devon retweeted
In-N-Out makes its own hamburger patties, in its own plants. The company buys whole chucks, takes the bones out, grinds the meat and forms the patties there. Then they're delivered to the restaurants on In-N-Out's own trucks. It opened its first patty plant in 1963 and now has plants in California, Texas and Colorado.
In-N-Out says its restaurants get fresh, never-frozen ingredients delivered throughout the week. The company has said for years that it won't open a store farther than it can truck fresh patties from its own plants. When it announced Tennessee in 2023, Lynsi Snyder said its meat plant in Lancaster, Texas, would supply patties to the Tennessee stores.
The fries are cut from whole potatoes in each restaurant. The menu board has 3 burgers, fries, shakes and drinks.
In-N-Out says its restaurants get fresh, never-frozen ingredients delivered throughout the week. The company has said for years that it won't open a store farther than it can truck fresh patties from its own plants. When it announced Tennessee in 2023, Lynsi Snyder said its meat plant in Lancaster, Texas, would supply patties to the Tennessee stores.
The fries are cut from whole potatoes in each restaurant. The menu board has 3 burgers, fries, shakes and drinks.
425
1,038
21,360
1,558,265
And here’s how most of my feed responded. I don’t know why but I always enjoy these debates.
I suggest Richard Sutton read this great article called “The Bitter Lesson” which explains why more computation worked to scale intelligence ☺️ incompleteideas.net/IncIdeas/Bitte…
0
0
0
137
There’s been some debate after @RichardSSutton’s tweet yesterday, mostly disagreeing with him, at least on my feed. Alex goes into some detail in his defense.
The bitter lesson has always been not just a push to leverage computation but a critique on methods. Methods that generalize and scale better win, and modeling those methods after human intuition is a waste of time. Somehow, much of the world forgot that methods matter.
Many of our methods today do not generalize or scale well. The results:
- tapping out the world’s resources extremely early in the deployment of AI
- increasing intelligence requires scaling model size faster than compute efficiency can offset it
- humans are the primary bottleneck in AI deployment
- demand for human-curation of knowledge is so extreme that any company with access to a human workforce can quickly generate millions in data revenue
- robotics intelligence is limited and task-specific
The bitter lesson requires us to rethink our methods so that they better leverage computation and generalize, and it explicitly shows an evolution of algorithms over time. Believing that the bitter lesson expects us to indefinitely scale a single static method is incredibly unintellectual.
Here are some methods that need attention:
- training and inference without quadratic compute complexity
- removing the KV cache
- something better than massive pre-training with BCE only
- is backpropagation still the best learning method?
- in-context learning for robotics with massive memory efficiency improvements
- algorithm-hardware co-development
- how do we remove humans from harness development?
0
0
1
137
Joe Devon retweeted
I recently made the tough decision to leave OpenAI.
It was truly a privilege to make some small contributions to the company’s incredibly massive ambition. I will be forever grateful for @embirico convincing me to join and @thsottiaux’s belief in me.
The people at OpenAI were the hardest part to leave behind. I am grateful to have worked directly with literally hundreds and hundreds of the smartest people I’ve ever met — across product, engineering, and GTM. Many of the best people I know also decided to join us during my time there, further proving how magnetic the culture is.
Even though I was there for a short time, it feels like the company went through 3 or 4 evolutions. I’m incredibly bullish on OpenAI because it is able to swarm on the hardest problems, move incredibly quickly, and adapt to all the craziness in the world today. I’ve always said, it’s a very non-regrettable choice to work at OpenAI, and that is more true than ever :)
Im pretty bad at taking breaks, but I look forward to grinding out a couple days on Maplestory Classic and then getting back to work 🚀
It was truly a privilege to make some small contributions to the company’s incredibly massive ambition. I will be forever grateful for @embirico convincing me to join and @thsottiaux’s belief in me.
The people at OpenAI were the hardest part to leave behind. I am grateful to have worked directly with literally hundreds and hundreds of the smartest people I’ve ever met — across product, engineering, and GTM. Many of the best people I know also decided to join us during my time there, further proving how magnetic the culture is.
Even though I was there for a short time, it feels like the company went through 3 or 4 evolutions. I’m incredibly bullish on OpenAI because it is able to swarm on the hardest problems, move incredibly quickly, and adapt to all the craziness in the world today. I’ve always said, it’s a very non-regrettable choice to work at OpenAI, and that is more true than ever :)
Im pretty bad at taking breaks, but I look forward to grinding out a couple days on Maplestory Classic and then getting back to work 🚀
125
41
1,630
316,101
(I didn't realize you were on X @LijanaWallenda)
Ever fallen down in life and found it hard to get up?
I have. Recently. It's hard.
But not nearly as hard as what my @NSALasVegas colleague Lijana has done.
Note Lijana's famous last name. Wallenda. The family known around the world for their dangerous high wire acts.
Lijana had a tragic accident performing a pyramid high wire act. After intensive rehab, she came back and crossed Times Square on a high wire on ABC with her brother Nik. I get scared just thinking about it.
I was mesmerized talking to her after this presentation.
She has the will, heart and guts of a Champion.
If you ever need an inspirational speaker for your conference, ask Lijana.
Ever fallen down in life and found it hard to get up?
I have. Recently. It's hard.
But not nearly as hard as what my @NSALasVegas colleague Lijana has done.
Note Lijana's famous last name. Wallenda. The family known around the world for their dangerous high wire acts.
Lijana had a tragic accident performing a pyramid high wire act. After intensive rehab, she came back and crossed Times Square on a high wire on ABC with her brother Nik. I get scared just thinking about it.
I was mesmerized talking to her after this presentation.
She has the will, heart and guts of a Champion.
If you ever need an inspirational speaker for your conference, ask Lijana.
Had the privilege of sharing a little bit of my story at the last @nsalasvegas meeting.
Thank you, @samhornintrigue for bringing me up on stage and giving me the opportunity to share.
0
0
0
122
Omar and his collaborators keep doing such interesting work. From Astra editing Qwen open weights, which went viral the other day, to a new retriever that "updates its weights during search, using gradients derived at inference time from the judgments of a reasoning model"
Never felt closer to a fundamental change in how we build retrievers, and a different kind of memory for agents.
ARI is basically test-time training for retrievers, where gradients derived from a reasoning model's preferences teach the retriever where to look *at search time*.
Lots more to do here to get oblique search queries as manageable as non-oblique ones, and that will include first figuring out sub-linear (late-interaction-esque) search.
But the path ahead has never been clearer. Read @dianetc_'s blog!
0
0
3
228
Joe Devon retweeted
A lot of people ask how did I learn to fine tune legal models? The answer stems from my desire to learn how things work at their core, and since LLM's are a many layered onion the core is deep....and also makes you cry a lot.
The two factors that pushed me into peeling back those layers were: 1. the models at the time (o1, claude 3) left a lot to be desired on the legal side; 2. I read reports about how they created the models and realized to advise clients I need to understand how it works. Reading about a LoRA adapter is not the same as watching it's loss curve drop in front of your eyes.
So I learned about @UnslothAI and used their free notebooks. I remember training llama on a really common legal task with a dataset I'd assembled. The resutls came back and I was thrilled as everything seemed so easy. But I learned a very valuable lesson after sharing my excitement and feeling confident in its results......always run it against a hold out set. It had memorized the training data, so anything new it saw it was actually WORSE on than the base.
So that was the real lesson, and it had nothing to do with what I had access to. Training eventually becomes the easier part, and the more difficult part is knowing whether what you have is any good. People kind of skip that because the outputs look good to them.
Three things I’d tell anyone starting:
1. Split your data before you touch the training script and create the hold out set and put it in a different folder with a .md that clearly says what it is.
2. Write down what the outcome and output is supposed to be first.
3. Small models memorize fast. If your loss drops to near zero, be suspicious, not excited.
That failure is why I now spend as much time on evaluation rubrics as on training itself. Also, people today will never know the pain of having to deal with pytorch version hell (I literally spent 4 days on runpod trying to figure out versioning).
If you’re a lawyer curious about this: start. The barrier is lower than you think. Just don’t trust your first result.
The two factors that pushed me into peeling back those layers were: 1. the models at the time (o1, claude 3) left a lot to be desired on the legal side; 2. I read reports about how they created the models and realized to advise clients I need to understand how it works. Reading about a LoRA adapter is not the same as watching it's loss curve drop in front of your eyes.
So I learned about @UnslothAI and used their free notebooks. I remember training llama on a really common legal task with a dataset I'd assembled. The resutls came back and I was thrilled as everything seemed so easy. But I learned a very valuable lesson after sharing my excitement and feeling confident in its results......always run it against a hold out set. It had memorized the training data, so anything new it saw it was actually WORSE on than the base.
So that was the real lesson, and it had nothing to do with what I had access to. Training eventually becomes the easier part, and the more difficult part is knowing whether what you have is any good. People kind of skip that because the outputs look good to them.
Three things I’d tell anyone starting:
1. Split your data before you touch the training script and create the hold out set and put it in a different folder with a .md that clearly says what it is.
2. Write down what the outcome and output is supposed to be first.
3. Small models memorize fast. If your loss drops to near zero, be suspicious, not excited.
That failure is why I now spend as much time on evaluation rubrics as on training itself. Also, people today will never know the pain of having to deal with pytorch version hell (I literally spent 4 days on runpod trying to figure out versioning).
If you’re a lawyer curious about this: start. The barrier is lower than you think. Just don’t trust your first result.
6
25
309
20,801
The father of that brave 20 year old LearJet copilot gave his first interview. He listened to that heartbreaking audio several times.
And he rightly came away proud of her for facing her fate stoically and reporting to the ATC all the way to the end. He noted that hardcore military types often aren't as brave in similar situations and she was rock solid.
I'm glad that at least the family can be so proud of her. Godspeed to the families. (And sorry there are some idiots on social media posting hurtful and wrong information.)
And he rightly came away proud of her for facing her fate stoically and reporting to the ATC all the way to the end. He noted that hardcore military types often aren't as brave in similar situations and she was rock solid.
I'm glad that at least the family can be so proud of her. Godspeed to the families. (And sorry there are some idiots on social media posting hurtful and wrong information.)
1
0
7
355
Joe Devon retweeted
A big part of the idea behind third party evaluation/auditing is that the identity and biases of an org doing the analysis are important, otherwise we’d take the companies’ word for it or require full transparency.
Even without full info, the METR report is very notable because of METR’s clear track record of, if anything, underestimating AI progress, not overestimating it.
Even without full info, the METR report is very notable because of METR’s clear track record of, if anything, underestimating AI progress, not overestimating it.
7
4
59
4,744
Very interesting thoughts from Dario. In the dust of Canva's challenges, I can't help but wonder about what Alex Karp said about enterprise companies being scared of training Anthropic and OpenAI models on their proprietary data and turning them into competitors. Just like the Amazon business model in retail.
I do still have a Canva Pro account, but I never use it because Fable does all my design work now.
How many enterprises are paying Anthropic big bucks today and competing with them for clients tomorrow?
I do still have a Canva Pro account, but I never use it because Fable does all my design work now.
How many enterprises are paying Anthropic big bucks today and competing with them for clients tomorrow?
Dario Amodei: A lot of AI generated videos and images are fake, addictive, and just slop. Anthropic isn't interested in competing in that market.
“Google and OpenAI are fighting it out in consumer. It's existential to both of them.”
“It's existential to OpenAI because that's xistential to Google because they have the search business.”
“For enterprise business, there's not really a demand for cats riding donkeys or whatever consumer video people want.”
“If we ever need it, we can just contract a model from someone else.”
“A lot of short-form video is fake, a lot of it's pretty addictive, a lot of it's slop.”
“It's not a part of the market that I'm tripping over myself to get involved in.”
0
0
3
400
Joe Devon retweeted
Today Physical Superintelligence PBC is announcing a proof of the Aaronson-Ambainis conjecture, one of the central open problems in quantum complexity, with major implications for quantum computing speedup.
psi.inc/research/aaron…
Thread 👇
psi.inc/research/aaron…
Thread 👇
165
669
4,048
1,188,349
Based on the quality of Armin and Mario alone, this is a great company to be a part of
We are looking for teammates who operate with intensity, integrity, and have a yearning to craft software and open protocols that strengthen human agency with us
earendil.com/join/
0
0
0
116
Joe Devon retweeted
Loved today’s Local AI Sessions with @javierbits and @BHIIKTOR. I had fun listening. I didn’t get a chance to speak, but there were a few things I wanted to add.
I’m a Mac user and use MLX. oMLX by Jun Kim (@jundotkim) has been great for me. I tend to get more tokens per second with it on my setup.
Its prefix cache can reuse matching prompt context instead of processing it again, which can make an answer start sooner. That’s separate from how quickly it generates new tokens once it starts.
I also like LM Studio for discovering models and checking what will fit my hardware. For coding, I sometimes use LM Studio Bionic with the dense Qwen 27B model.
I love Gemma and LFM too. I also run models on my iPhone 18 Pro. I love our app Perspective Intelligence, built by @mikedoise, for on-device AI.
What local models are you using, and what are you running them on?
Listen here: x.com/i/spaces/1AMXd…
#LocalAI #MLX
I’m a Mac user and use MLX. oMLX by Jun Kim (@jundotkim) has been great for me. I tend to get more tokens per second with it on my setup.
Its prefix cache can reuse matching prompt context instead of processing it again, which can make an answer start sooner. That’s separate from how quickly it generates new tokens once it starts.
I also like LM Studio for discovering models and checking what will fit my hardware. For coding, I sometimes use LM Studio Bionic with the dense Qwen 27B model.
I love Gemma and LFM too. I also run models on my iPhone 18 Pro. I love our app Perspective Intelligence, built by @mikedoise, for on-device AI.
What local models are you using, and what are you running them on?
Listen here: x.com/i/spaces/1AMXd…
#LocalAI #MLX
6
1
6
440