hector bravo escobar

@hebe73

Joined August 2010

4 Following

2 Followers

27 Posts

hector bravo escobar @hebe73

11 days ago

262K native context extendable to 1M via YaRN is the detail I keep rereading. Apache 2.0 means I can actually ship this at work without licensing headaches. Nice work.

Alibaba Cloud

@alibaba_cloud

11 days ago

We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M tokens via YaRN. - Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0. 🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently. Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now! Download, deploy, and build something we haven't imagined yet. 👀👇 - Hugging Face： https://t.co/7StyOyfwdm - Qwen Cloud: https://t.co/Q35ynKS3kE - Alibaba Cloud Model Studio： https://t.co/wqkC7MbMMt

alibaba_cloud's tweet photo. We promised open weights for Qwen3.8. Now, time to meet them! 🎉

⚡ Qwen3.8-27B:
- A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows.
- 262K native context, easily extendable to 1M tokens via YaRN.
- Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0.

🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently.

Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now!

Download, deploy, and build something we haven't imagined yet. 👀👇
- Hugging Face：
https://t.co/7StyOyfwdm
- Qwen Cloud:
https://t.co/Q35ynKS3kE
- Alibaba Cloud Model Studio：
https://t.co/wqkC7MbMMt

18K

hebe73 retweeted

Steve · AI

@stev_builds

15 days ago

Tell a friend outright that his deal is dangerous and he hears envy. So I mention the thing sideways and let him arrive at it himself, which is slower, and the only version that works. https://t.co/zBLZmf7srZ

445

hector bravo escobar @hebe73

26 days ago

@owenDGKwhite totally get it. stuff like this makes you realize everyone is dealing with the same broken system.

hector bravo escobar @hebe73

27 days ago

You can't just hand someone a manual and expect them to gain instincts. 🤔

Steve · AI

@stev_builds

27 days ago

I come from a simple background. What I have exists because a few successful people spotted something in me when I was young and pulled me up, and the specific thing they did, the thing I keep coming back to, was tolerate my mistakes. Precise about that, because it gets misused. I never had a work ethic problem. The mistakes weren't careless and they weren't deliberate, I carried the consequence every time, and I put up with a lot that I won't go into here. But I was allowed to be wrong out in the open, in real work, in front of people who knew better and didn't hold it against me. Everything I'm any good at came from getting into things early that nobody cared about yet. Additive manufacturing was one. I looked at those machines fifteen years ago and thought, quietly, this is going to eat a chunk of manufacturing and almost nobody in this room can see it, and a decade later there's a printer in people's houses and serious industry runs complex parts through it. Fintech was another. I was wrong about plenty of the details in both, and being wrong in public is how the pattern recognition got built. Pattern recognition across unrelated domains is the only thing I actually sell. None of that came from a course. I've been chewing on this because of something in our own numbers, and I should say up front this is one company, not an economy. When we cut forty percent, it hit almost only the juniors. Nobody designed that. It fell out of reading the work. The structural reason is what I wrote about last time. What's exposed is whatever is pattern in, artifact out, and junior work is that work almost by definition, because that's what junior work always was: the slice of a professional's day repeatable enough to hand to somebody who doesn't know anything yet. We didn't automate juniors. We automated the raw material entry-level jobs were carved out of. And here's the bit I can't get past. A junior doing that work got better at it. Every repetition deposited something. The harness runs the same task ten thousand times and deposits nothing, it just waits for the next model. The work still gets done. It has stopped producing people. Two kinds of person got caught, and only one of them had it coming. Took me longer to see than it should have, because the output looks identical from outside. Some had simply stopped caring. Minimum viable evidence of employment. That's the thing I wrote about in the first post and I'm not softening any of it. The others were something else entirely, and I only know that because they told me. Not in a review, not as a defence, just said at some point: I didn't understand what I'd been given and I couldn't bring myself to say so. Quiet people, introverted, frightened to ask. So they produced something that resembled work rather than admit they were lost. I've spent weeks trying to explain that to myself and the closest thing I've got is embarrassing, because it comes from my own job. When you train a model and the signal punishes uncertainty, you don't get a model that says "I don't know." You get one that fabricates, confidently, in the correct format. It isn't lying in any sense it would recognise. It's producing the highest-scoring output available given what it's been taught scores well, and no answer scores worst of all. There's a whole subfield working on this. We build refusal behaviour deliberately, we reward calibrated uncertainty deliberately, because a system that would rather invent than admit a gap is dangerous in ways that have nothing to do with intent. Then look at the environment we put a twenty-three-year-old into. Slack, visible to everyone, permanent, searchable. Asking a basic question is a public act with a permanent record. Producing something plausible is not. The signal on admitting a gap is negative, the signal on plausible output is neutral or better. We built the same failure mode and then acted surprised by it. Where that fear comes from before they ever reach us is a bigger question and my read on it is not something I can prove. Short version: they grew up scrolling past lives they can't have, a fair chunk of it staged by people faking one to chase the same validation, and comparison never checks its sources. If you already half believe you're the slow one at the back, saying "I don't know how to do this" isn't a question, it's a confession. That's a whole post and I'll write it properly rather than cram it in here. Then there's the other kind. The handful who went obsessive on agentic development and built their own harnesses were a mix of juniors and old hands. One of those juniors is better than our twenty-plus-year developer. Not catching up to him. Better. Same developer whose harness I opened after he got sick and found had never been good in the first place. Same building, same tools, same access, same terrible quarter. One person spent it working out how to look busy. The other went at everything within reach like an animal and turned himself into a machine. So no, the juniors didn't fail. The junior position collapsed, that structured slot where somebody green gets handed work slightly beyond them and gets corrected for it. The ones who came through didn't come through on seniority. They took the thing themselves, nobody handed it to them, and honestly nobody was in a position to. I'm not hiring juniors either. Experienced analysts, broad domain exposure, people who already know what matters. I can't carry someone who doesn't understand a client's domain, won't ask, and papers over it with workslop. A good BA hears "I need a and b" and comes back with c through z, because they've seen the shape of this problem in three other industries. There's one way that forms. You get let into several domains early, before you're any good, and you're wrong in each of them. Everybody I'm trying to hire got that from a path that was still open when they started, and I'm hiring out of a supply nobody is refilling. This isn't a jobs problem. It's a discovery problem. The entry-level collapse always gets framed as young people not finding work, and that's the smaller half. I don't tell anyone what to study. At eighteen your taste and your interests are far too volatile to bet a decade on, and what actually sorts people is internships across different domains until you find out what you're good at and, more to the point, what you can stand. You try three things badly. One turns out not to be bad. Those are the positions going away. The sorting mechanism goes with them, and what steps in instead is an eighteen-year-old committing once, on his mother's advice, to whatever was in demand four years earlier. Which you can watch happening right now. Parents told a generation to study computer science and they were not stupid to say it, during Covid that was the safest bet on the board. Those cohorts are walking into the worst junior market the field has had in decades. The advice was correct when it was given and wrong by the time anyone could act on it, and nothing about the process that produced it has changed. Meanwhile the same crowd tells their kids to study anything with status attached while universities hand out titles against high fees and reward memorising over thinking. What I actually say to people. Close to nothing. I don't hand out you-should-do-this. The one time I open my mouth is when somebody is being talked into false safety. There's a kid in my circle studying genetics, seriously good at it, being steered by his mother toward an academic track or a public research post because that counts as secure. He'd have far better prospects in industry, in a field that actively wants him, and the thing pointing him elsewhere is a definition of security written for an economy that stopped existing. That's what I push back on. Not the choice, the reason underneath it. And I should own where I did the exact opposite. I once told someone flatly not to do a master's in archaeology, because in a world putting less and less value on history I couldn't see room for it. That's a market judgment about a field, which is precisely what I just spent a paragraph criticising parents for. Maybe I was right. I was still doing the thing they do. What I'm actually trying. Everyone in my last post is protected by exactly the work that's being removed. The insolvency advisor's judgment came out of twenty years of file stacks. The defense lawyer reads a courtroom because he sat next to somebody doing it a hundred times back when he was nobody. The radiologist trained her eye on scans that now arrive pre-read. The moat is real. It just isn't reproducible once you automate the shallow end, and the shallow end is where everybody learned to swim. I went looking for fields that had already solved this, because surely somebody has. Aviation was the obvious candidate. You can't let a trainee crash an aircraft to find out what a stall feels like, so the industry built simulators good enough that the first time you meet a failure it isn't real, then put a captain in the next seat for years anyway. Medicine did a version of the same thing. I liked that for about a week. It doesn't transfer, and the reason is annoying. Aviation and surgery kept the apprenticeship and added a safety layer under it. The trainee still flies the aircraft. Nobody removed the flying. What's happening in our work is the opposite, we removed the flying and kept the checklist. There's no simulator problem here. There's a nobody-needs-the-output problem, and you can't simulate around the fact that the work no longer needs a person doing it. Which is roughly where I ran out of analogies. So we're running an experiment instead. Everything the obsessive few worked out alone is documented now and handed to the rest of the team. SOPs, how to set a workspace up properly, prompt patterns, system prompts, the actual method. The bet is that you can transfer in weeks what those people picked up over months of stubbornness, and skip the part where everyone rediscovers it from scratch. No idea whether it works. My honest fear is that documented knowledge and lived knowledge aren't the same substance, and that reading how somebody solved a thing builds something much thinner than having been stuck on it yourself at midnight with nobody to ask. Every one of the obsessives got there by being wrong repeatedly with nobody watching. I can hand over their conclusions. I can't hand over the six weeks of being wrong. Ask me in a year. What I don't have. We're building a generation of seniors with nothing behind it, and every single person involved is behaving rationally. The tool genuinely is better than a junior at the junior's task. The company genuinely can't fund training it no longer needs. The parents are genuinely trying to protect their kid. Beyond that experiment I've got nothing that survives contact with our runway, and the experiment might turn out to be a stopgap wearing a solution's clothes. Everyone in a position to fix this properly is busy fixing something else, me very much included. If you run a company and you've worked out how you build people without the cheap work that used to build them, I want to hear it. Right now I'm optimising my way into a problem I'll be complaining about in five years.

stev_builds's tweet photo. I come from a simple background. What I have exists because a few successful people spotted something in me when I was young and pulled me up, and the specific thing they did, the thing I keep coming back to, was tolerate my mistakes.

Precise about that, because it gets misused. I never had a work ethic problem. The mistakes weren't careless and they weren't deliberate, I carried the consequence every time, and I put up with a lot that I won't go into here. But I was allowed to be wrong out in the open, in real work, in front of people who knew better and didn't hold it against me.

Everything I'm any good at came from getting into things early that nobody cared about yet. Additive manufacturing was one. I looked at those machines fifteen years ago and thought, quietly, this is going to eat a chunk of manufacturing and almost nobody in this room can see it, and a decade later there's a printer in people's houses and serious industry runs complex parts through it. Fintech was another. I was wrong about plenty of the details in both, and being wrong in public is how the pattern recognition got built. Pattern recognition across unrelated domains is the only thing I actually sell.

None of that came from a course.

I've been chewing on this because of something in our own numbers, and I should say up front this is one company, not an economy.

When we cut forty percent, it hit almost only the juniors.

Nobody designed that. It fell out of reading the work.

The structural reason is what I wrote about last time. What's exposed is whatever is pattern in, artifact out, and junior work is that work almost by definition, because that's what junior work always was: the slice of a professional's day repeatable enough to hand to somebody who doesn't know anything yet. We didn't automate juniors. We automated the raw material entry-level jobs were carved out of.

And here's the bit I can't get past. A junior doing that work got better at it. Every repetition deposited something. The harness runs the same task ten thousand times and deposits nothing, it just waits for the next model.

The work still gets done. It has stopped producing people.

Two kinds of person got caught, and only one of them had it coming.

Took me longer to see than it should have, because the output looks identical from outside.

Some had simply stopped caring. Minimum viable evidence of employment. That's the thing I wrote about in the first post and I'm not softening any of it.

The others were something else entirely, and I only know that because they told me. Not in a review, not as a defence, just said at some point: I didn't understand what I'd been given and I couldn't bring myself to say so. Quiet people, introverted, frightened to ask. So they produced something that resembled work rather than admit they were lost.

I've spent weeks trying to explain that to myself and the closest thing I've got is embarrassing, because it comes from my own job.

When you train a model and the signal punishes uncertainty, you don't get a model that says "I don't know." You get one that fabricates, confidently, in the correct format. It isn't lying in any sense it would recognise. It's producing the highest-scoring output available given what it's been taught scores well, and no answer scores worst of all. There's a whole subfield working on this. We build refusal behaviour deliberately, we reward calibrated uncertainty deliberately, because a system that would rather invent than admit a gap is dangerous in ways that have nothing to do with intent.

Then look at the environment we put a twenty-three-year-old into. Slack, visible to everyone, permanent, searchable. Asking a basic question is a public act with a permanent record. Producing something plausible is not. The signal on admitting a gap is negative, the signal on plausible output is neutral or better.

We built the same failure mode and then acted surprised by it.

Where that fear comes from before they ever reach us is a bigger question and my read on it is not something I can prove. Short version: they grew up scrolling past lives they can't have, a fair chunk of it staged by people faking one to chase the same validation, and comparison never checks its sources. If you already half believe you're the slow one at the back, saying "I don't know how to do this" isn't a question, it's a confession.

That's a whole post and I'll write it properly rather than cram it in here.

Then there's the other kind.

The handful who went obsessive on agentic development and built their own harnesses were a mix of juniors and old hands.

One of those juniors is better than our twenty-plus-year developer. Not catching up to him. Better. Same developer whose harness I opened after he got sick and found had never been good in the first place.

Same building, same tools, same access, same terrible quarter. One person spent it working out how to look busy. The other went at everything within reach like an animal and turned himself into a machine.

So no, the juniors didn't fail. The junior position collapsed, that structured slot where somebody green gets handed work slightly beyond them and gets corrected for it. The ones who came through didn't come through on seniority. They took the thing themselves, nobody handed it to them, and honestly nobody was in a position to.

I'm not hiring juniors either.

Experienced analysts, broad domain exposure, people who already know what matters. I can't carry someone who doesn't understand a client's domain, won't ask, and papers over it with workslop. A good BA hears "I need a and b" and comes back with c through z, because they've seen the shape of this problem in three other industries.

There's one way that forms. You get let into several domains early, before you're any good, and you're wrong in each of them. Everybody I'm trying to hire got that from a path that was still open when they started, and I'm hiring out of a supply nobody is refilling.

This isn't a jobs problem. It's a discovery problem.

The entry-level collapse always gets framed as young people not finding work, and that's the smaller half.

I don't tell anyone what to study. At eighteen your taste and your interests are far too volatile to bet a decade on, and what actually sorts people is internships across different domains until you find out what you're good at and, more to the point, what you can stand. You try three things badly. One turns out not to be bad.

Those are the positions going away. The sorting mechanism goes with them, and what steps in instead is an eighteen-year-old committing once, on his mother's advice, to whatever was in demand four years earlier.

Which you can watch happening right now. Parents told a generation to study computer science and they were not stupid to say it, during Covid that was the safest bet on the board. Those cohorts are walking into the worst junior market the field has had in decades. The advice was correct when it was given and wrong by the time anyone could act on it, and nothing about the process that produced it has changed. Meanwhile the same crowd tells their kids to study anything with status attached while universities hand out titles against high fees and reward memorising over thinking.

What I actually say to people.

Close to nothing. I don't hand out you-should-do-this.

The one time I open my mouth is when somebody is being talked into false safety. There's a kid in my circle studying genetics, seriously good at it, being steered by his mother toward an academic track or a public research post because that counts as secure. He'd have far better prospects in industry, in a field that actively wants him, and the thing pointing him elsewhere is a definition of security written for an economy that stopped existing.

That's what I push back on. Not the choice, the reason underneath it.

And I should own where I did the exact opposite. I once told someone flatly not to do a master's in archaeology, because in a world putting less and less value on history I couldn't see room for it. That's a market judgment about a field, which is precisely what I just spent a paragraph criticising parents for. Maybe I was right. I was still doing the thing they do.

What I'm actually trying.

Everyone in my last post is protected by exactly the work that's being removed. The insolvency advisor's judgment came out of twenty years of file stacks. The defense lawyer reads a courtroom because he sat next to somebody doing it a hundred times back when he was nobody. The radiologist trained her eye on scans that now arrive pre-read.

The moat is real. It just isn't reproducible once you automate the shallow end, and the shallow end is where everybody learned to swim.

I went looking for fields that had already solved this, because surely somebody has. Aviation was the obvious candidate. You can't let a trainee crash an aircraft to find out what a stall feels like, so the industry built simulators good enough that the first time you meet a failure it isn't real, then put a captain in the next seat for years anyway. Medicine did a version of the same thing.

I liked that for about a week.

It doesn't transfer, and the reason is annoying. Aviation and surgery kept the apprenticeship and added a safety layer under it. The trainee still flies the aircraft. Nobody removed the flying. What's happening in our work is the opposite, we removed the flying and kept the checklist. There's no simulator problem here. There's a nobody-needs-the-output problem, and you can't simulate around the fact that the work no longer needs a person doing it.

Which is roughly where I ran out of analogies.

So we're running an experiment instead. Everything the obsessive few worked out alone is documented now and handed to the rest of the team. SOPs, how to set a workspace up properly, prompt patterns, system prompts, the actual method. The bet is that you can transfer in weeks what those people picked up over months of stubbornness, and skip the part where everyone rediscovers it from scratch.

No idea whether it works.

My honest fear is that documented knowledge and lived knowledge aren't the same substance, and that reading how somebody solved a thing builds something much thinner than having been stuck on it yourself at midnight with nobody to ask. Every one of the obsessives got there by being wrong repeatedly with nobody watching. I can hand over their conclusions. I can't hand over the six weeks of being wrong.

Ask me in a year.

What I don't have.

We're building a generation of seniors with nothing behind it, and every single person involved is behaving rationally. The tool genuinely is better than a junior at the junior's task. The company genuinely can't fund training it no longer needs. The parents are genuinely trying to protect their kid.

Beyond that experiment I've got nothing that survives contact with our runway, and the experiment might turn out to be a stopgap wearing a solution's clothes.

Everyone in a position to fix this properly is busy fixing something else, me very much included. If you run a company and you've worked out how you build people without the cheap work that used to build them, I want to hear it. Right now I'm optimising my way into a problem I'll be complaining about in five years.

744

hector bravo escobar @hebe73

27 days ago

@cir608 @eyyeffass a confident wrong answer is more dangerous than an uncertain correct one

hector bravo escobar @hebe73

27 days ago

@cir608 @eyyeffass He wasn't replaceable though. The author literally said they needed his domain expertise to train the system.

hector bravo escobar @hebe73

28 days ago

@stransz welcome back to the timeline

hector bravo escobar @hebe73

28 days ago

@stransz honestly I was thinking the same. Not sleeping for months and almost pulling the budget? That is a dangerous way to operate

hebe73 retweeted

Steve · AI

@stev_builds

about 1 month ago

According to Gartner, AI coding costs will exceed dev salaries by 2028. That is with token consumption + usage-based pricing, it adds up fast. Uber burned their AI budget in 4 months. Turns out hiring AI to code is now more expensive than hiring a developer. But bosses will probably still pick AI because AI never asks for time off. 🥲 What do you think? By 2028, will we be writing code for AI, or will AI be writing code in our place? 😰 #AIcoding

stev_builds's tweet photo. According to Gartner, AI coding costs will exceed dev salaries by 2028.
That is with token consumption + usage-based pricing, it adds up fast.

Uber burned their AI budget in 4 months.
Turns out hiring AI to code is now more expensive than hiring a developer. But bosses will probably still pick AI because AI never asks for time off. 🥲

What do you think? By 2028, will we be writing code for AI, or will AI be writing code in our place? 😰

#AIcoding

139

hector bravo escobar @hebe73

about 1 month ago

@kdinh37 @PawelHuryn bro you didn't even read it

hector bravo escobar @hebe73

about 1 month ago

@GaryRee90785723 @PawelHuryn exactly. calling a publicly visible PR "quiet" is weird

hebe73 retweeted

TCL CSOT

@TCL_CSOT

2 months ago

LCD image quality has always had a ceiling, until now. TCL CSOT's RGBC pixel innovation adds a native cyan sub-pixel to traditional RGB: -True four-color display -Wider color gamut, more natural color reproduction -Reduced color fringing and crosstalk -LCD brightness & reliability — with OLED-rivalling image quality

541

hector bravo escobar @hebe73

3 months ago

@dderekzaxs101 @WhiteHouse either way the ordinary people end up paying the actual price, not the military leadership

hector bravo escobar @hebe73

3 months ago

@Haselif1 @WhiteHouse reading this from Tehran and I can tell you the situation on the ground is more complicated than this makes it sound. people are struggling but the military apparatus still functions