Jason retweeted
Introducing Nestor!
In standard MoEs, tokens choose their FFN experts. What if they could choose their attention modules, too? 🤔
We introduce a transformer with nested sparsity for scalable pretraining and continual learning.
In standard MoEs, tokens choose their FFN experts. What if they could choose their attention modules, too? 🤔
We introduce a transformer with nested sparsity for scalable pretraining and continual learning.
9
54
356
48,521
Jason retweeted
How many pencils, how many humans, how many hours would it have taken to recreate the Adobe suite by hand?
Todos los productos de Adobe reimplementados desde cero, gratuitos y de código abierto
→ getartcraft.com/apps
325
360
8,317
590,186
Statistic teaches us how to mislead others. The chart shows the risk increasing without showing how many applications we made over the past few years. Tools like lovable will lead to tons of cyber risks while most application have 0 number or user. When we reset the "graph" to ratio graph you will see another story.
Cyber is having a moment
Across 21 major software companies, including Apple, AWS, Microsoft, and Google:
- Reported critical vulnerabilities never cleared 100 per month in four years
- Since spring they've jumped to over 600 per month
0
0
0
68
You don't have to tell the coding agent what kind of "techniques" to use cause it almost always use the proper one.
For example when building a "google doc", it will use web socket and if your terminal connect GitHub it will always do the CI/CD jobs.
I am not saying what we have learnt so far are useless, just like copilot we are useful when it's emergency or flying with fighter jet every tiny details are important. Yet most people and most project only need basic security and knowledge about computer science.
In this "AI era" and also true before the AI era what to build is way more important than how you build. It has always been the case. Just like 10 years ago, everyone calling meta code bases are shit, it doesn't seems having any correlation between revenue and "technical".
This brings a question for every vibe haters: we are crystal because we can generate the revenue or we know more technical skill set than others ?
For example when building a "google doc", it will use web socket and if your terminal connect GitHub it will always do the CI/CD jobs.
I am not saying what we have learnt so far are useless, just like copilot we are useful when it's emergency or flying with fighter jet every tiny details are important. Yet most people and most project only need basic security and knowledge about computer science.
In this "AI era" and also true before the AI era what to build is way more important than how you build. It has always been the case. Just like 10 years ago, everyone calling meta code bases are shit, it doesn't seems having any correlation between revenue and "technical".
This brings a question for every vibe haters: we are crystal because we can generate the revenue or we know more technical skill set than others ?
Do vibe coders actually understand:
Race conditions
Webhooks
Rate limits
Caching
MCP servers
CI/CD
Database indexes
Background jobs
Or are we just writing CRUD and hoping for the best? 😭
1
0
3
194
The first two companies boost their IPO valuation by threatening
@0xDevShah Anthropic 2026 expected run rate - ~$110B (24x)
OpenAI 2026 expected run rate - ~$70B+ (5.4x+)
Meta 2026 expected run rate - ~$295B (1.5x)
I'm so tired.
0
0
2
128
I am optimistic about AI development but frankly permissive about the valuation. LLM tasks are easily verifiable, while most tasks like writing or socializing are non-deterministic. Spending on AI for work will increase, but not for general public daily use, at least not for now in most cases.
98% of US households aren't paying for AI yet
More charts in State of Markets II: a16z.news/p/state-of-mar…
0
0
1
55
