elias tazartes 🥕🧑🌾 retweeted
totally agree minimizing human attention is the key unlock now and moving forward
this view from calvin on workflow changes correlating with productivity gain is also quite interesting - i think mine was similar
my step function gains came from:
- overnight loops (gnhf)
- idation and risk assessment (no-mistakes)
- interactive artifacts for planning (lavish-axi)
- use an agent to orchestrate agents (firstmate)
- remote ssh setup (sshhip)
this view from calvin on workflow changes correlating with productivity gain is also quite interesting - i think mine was similar
my step function gains came from:
- overnight loops (gnhf)
- idation and risk assessment (no-mistakes)
- interactive artifacts for planning (lavish-axi)
- use an agent to orchestrate agents (firstmate)
- remote ssh setup (sshhip)
Sharing an update on my agentic workflow!
Today, there are only two REAL bottlenecks left in software development. Compute and Human Attention. Not PRs, Planning, CI/CD. Those aren't real bottlenecks. You only think they are because they consume your Human Attention.
You can solve the Compute problem by throwing money at it. You can solve the Human Attention problem by throwing Compute at it.
I reframed my workflow optimization around minimizing places that need my attention (and continue to do so). This blog post talks about my new setup.
I hope...
open.substack.com/pub/thebuildin…
27
28
585
40,993
elias tazartes 🥕🧑🌾 retweeted
they’re calling it the black October lmao.
> node.js: security vulnerability.
> PHP: critical RCE vulnerabilities.
> tensorlake: npm package infected.
> liteLLM: malicious PyPI + Docker releases
> Railway: customer secrets exposed
> zeabur - infrastructure breached.
> OpenCode: AI API keysI: official binaries backdoored.
> RankMath PRO: malware inside official plugin.
> TeamUpdraft: customer data compromised.
> ConnectWise: cloud backup identities compromised.
> SubQuery: backdoored npm package.
> GhostAction: GitHub repos with malicious workflows (related to the broader Shai-Hulud/ChainDrop activity).
> eMudhra: certificate authority keys compromised (with downstream impact on Tailscale).
this some complete madness bro. 1/3 of the internet is fucked.
> node.js: security vulnerability.
> PHP: critical RCE vulnerabilities.
> tensorlake: npm package infected.
> liteLLM: malicious PyPI + Docker releases
> Railway: customer secrets exposed
> zeabur - infrastructure breached.
> OpenCode: AI API keysI: official binaries backdoored.
> RankMath PRO: malware inside official plugin.
> TeamUpdraft: customer data compromised.
> ConnectWise: cloud backup identities compromised.
> SubQuery: backdoored npm package.
> GhostAction: GitHub repos with malicious workflows (related to the broader Shai-Hulud/ChainDrop activity).
> eMudhra: certificate authority keys compromised (with downstream impact on Tailscale).
this some complete madness bro. 1/3 of the internet is fucked.
133
584
3,793
119,780
elias tazartes 🥕🧑🌾 retweeted
🚨CLOSED SOURCE SOFTWARE IS DEAD
... especially in web3. Security researchers are now achieving hash-perfect recompilation of Rust projects' source code, using AI. And then they find Critical vulnerabilities.
Insanely cool work by @osec_io, love it👏
osec.io/blog/goonfi-de…
... especially in web3. Security researchers are now achieving hash-perfect recompilation of Rust projects' source code, using AI. And then they find Critical vulnerabilities.
Insanely cool work by @osec_io, love it👏
osec.io/blog/goonfi-de…
10
22
245
10,453
elias tazartes 🥕🧑🌾 retweeted
one of us
Linus Torvalds to the anti-AI crowd: stop using compilers if you want to prove they’re REAL programmers. LOL.
“you had better not use an editor to write that commit message either.
The kind of expectations *I* have is that people do their best and I expect people to understand ut that understanding has very little to do with what tools they use.
And quite honestly, in my experience, the people I respect and call intelligent are not the ones that shy away from using new and interesting tools.”
12
16
616
28,537
elias tazartes 🥕🧑🌾 retweeted
working on a product that reaches 100M TVL so fast is fucking amazing
A confidentiality milestone for onchain credit
$100M+ is now earning yield from Morpho Vaults via @zama, with positions and earnings kept in confidence.
1
3
50
991
elias tazartes 🥕🧑🌾 retweeted
I can now publish some results so I'm starting with our Verity FV bench. It is somewhat niche, which has the advantage of measuring the models' on kind of problem that is not part of their training set
github.com/lfglabs-dev/et…
github.com/lfglabs-dev/et…
Thanks to the kind people at @GoogleDeepMind I got early access to Gemini 4 for the past week. So far it's the best model I've ever used:
- excellent at formal verification
- very consistent
Also the first model whose writing doesn't seem "ai"
4
4
19
1,319
elias tazartes 🥕🧑🌾 retweeted
js was built in 10 days, what's the flex /s
Some rock solid engineers I know are getting REALLY ambitious in building with AI - building systems in ~months that would have been madness to consider before.
I’ll have a specific case study to share in 3-4 weeks about building a brand new language and runtime in months 😱
0
1
20
2,222
elias tazartes 🥕🧑🌾 retweeted
There seem to be two groups:
1. Better-is-better group believes that agents finally make the better-is-better solutions happen. Formal methods, static types, Nix, ...
2. Worse-is-better group believes agents will make worse-is-better solution happen because with agents they're no longer worse. Dynamic type systems, bash scripts over neat DSLs, agents running tests dynamically, ...
Quite funny to see
1. Better-is-better group believes that agents finally make the better-is-better solutions happen. Formal methods, static types, Nix, ...
2. Worse-is-better group believes agents will make worse-is-better solution happen because with agents they're no longer worse. Dynamic type systems, bash scripts over neat DSLs, agents running tests dynamically, ...
Quite funny to see
34
22
312
22,965
elias tazartes 🥕🧑🌾 retweeted
The setup I want:
- Always-on Mac at home
- Coding agent with computer use
- MacBook that syncs my sessions
- iOS app + voice agent paired with my Mac
So I can voice control everything on my machine through the mobile app
And work on those same session when I’m at my desk
- Always-on Mac at home
- Coding agent with computer use
- MacBook that syncs my sessions
- iOS app + voice agent paired with my Mac
So I can voice control everything on my machine through the mobile app
And work on those same session when I’m at my desk
239
35
1,516
169,820
elias tazartes 🥕🧑🌾 retweeted
i feel like swe interviews could probably collapse into just 2 technical rounds:
• system design (how well can you articulate your ideas, do you actually know how to engineer something)
• onsite project to build a real thing with agents (can you translate your intent into high quality outcomes)
everything else we used to ask is no longer necessary
• system design (how well can you articulate your ideas, do you actually know how to engineer something)
• onsite project to build a real thing with agents (can you translate your intent into high quality outcomes)
everything else we used to ask is no longer necessary
🗞️ if your team is too busy doing their 'normal job' to experiment with AI, you're preparing them to be replaced
ghuntley.com/replaced/
In this post, I guess this's the last time I'll say it. AI use is no longer optional if you wish to be employed. What follows is advice for iring managers on how to hire in 2026 to source AI-first candidates.
292
413
6,456
424,523
elias tazartes 🥕🧑🌾 retweeted
A few people yesterday asked how AI could impact FHE / @zama, so here is a quick tldr:
1. There is no evidence that lattices / FHE can be broken by AI, SI or quantum. If anything, there is a growing consensus that fully breaking lattices might imply P = NP.
2. However, it's almost certain that better attacks will be discovered, which would require increasing parameter size and taking a short-term hit on performance. The Zama protocol is modular, so we can increase parameters size or swap out any component quickly, without changing the UX or rotating user keys.
3. Zama uses FHE only for privacy, so breaking FHE would not allow stealing funds that are shielded, it would only allow seeing balances and amounts. This is unlike ZK, where breaking the cryptography means you can steal funds and mint infinite tokens.
4. We do a lot of work on cryptanalysis and improving the security of FHE. We spend millions a year on audits (AI and human), research grants, bounties, have 30 PhDs in the team, etc. We have a plan A (the current thing we do), but also a plan B and a plan C if things go wrong. Whatever the future throws at us, we will have a way to adapt to offer the same service.
5. Hashes have been broken in the past, many times (MD5, SHA1, etc), so saying they are more secure than lattices, which have been extensively studied for decades is a stretch to say the least. Not saying its impossible, but history is on the side of lattices here.
Lets also remember that if indeed lattices and elliptic curves are broken and there is no public-key encryption anymore, then hiding your balances onchain is the least of your worries, because the world as we know it would be over: no private communications, no online banking, your AI prompts would be public, etc.
1. There is no evidence that lattices / FHE can be broken by AI, SI or quantum. If anything, there is a growing consensus that fully breaking lattices might imply P = NP.
2. However, it's almost certain that better attacks will be discovered, which would require increasing parameter size and taking a short-term hit on performance. The Zama protocol is modular, so we can increase parameters size or swap out any component quickly, without changing the UX or rotating user keys.
3. Zama uses FHE only for privacy, so breaking FHE would not allow stealing funds that are shielded, it would only allow seeing balances and amounts. This is unlike ZK, where breaking the cryptography means you can steal funds and mint infinite tokens.
4. We do a lot of work on cryptanalysis and improving the security of FHE. We spend millions a year on audits (AI and human), research grants, bounties, have 30 PhDs in the team, etc. We have a plan A (the current thing we do), but also a plan B and a plan C if things go wrong. Whatever the future throws at us, we will have a way to adapt to offer the same service.
5. Hashes have been broken in the past, many times (MD5, SHA1, etc), so saying they are more secure than lattices, which have been extensively studied for decades is a stretch to say the least. Not saying its impossible, but history is on the side of lattices here.
Lets also remember that if indeed lattices and elliptic curves are broken and there is no public-key encryption anymore, then hiding your balances onchain is the least of your worries, because the world as we know it would be over: no private communications, no online banking, your AI prompts would be public, etc.
34
52
233
20,565
elias tazartes 🥕🧑🌾 retweeted
Mistral Large 4 was trained on 4,000 GPUs and Astra was trained on 100,000 of them.
We'll see what the kitten can pull off, but the disparity is pretty stark
We'll see what the kitten can pull off, but the disparity is pretty stark
Frontier scale, built in Europe.
Nearly 4,000 NVIDIA Grace Blackwell Superchips powered the training of Mistral Large 4, now in public preview.
Congrats to the @MistralAI team 🙌
56
112
2,953
168,027
elias tazartes 🥕🧑🌾 retweeted
The Zama team is kicking off @token2049 week in Singapore 🇸🇬
We're co-hosting @Yield_Summit today: a full day of keynotes, panels and roundtables on confidential onchain finance.
Grab a copy of the latest Privacy Paper, or just say hi if you're around.
We're co-hosting @Yield_Summit today: a full day of keynotes, panels and roundtables on confidential onchain finance.
Grab a copy of the latest Privacy Paper, or just say hi if you're around.
21
30
179
12,046
elias tazartes 🥕🧑🌾 retweeted
Excellent agentic coding interview.
Key points:
Build a “trust ladder”. The more you can constrain the agent’s output, the more you can trust it.
Instead of a mere “software factory”, try to create a “Michelin restaurant”. Create processes and tools that make beautiful code automatically.
Think about code reviews like sampling. You can’t afford to taste every dish, but you should spot check.
Ask “where am I the bottleneck? How can I get the agent to answer its own questions?”
Create inner loops and outer loops. Inner loops accomplish tasks. Outer loops coordinate inner loops and analyze the big picture.
Your skills should get smaller over time as agents improve.. Focus on workflows, not implementation details.
Connectors trigger new work. Connect your bots to Jira, GitHub issues, etc so they begin work automatically.
Make it easy for agents to verify their work. Create a CLI for your app and a skill that calls the CLI. Then agents can verify their work consistently and efficiently.
A good code base is easy to change.
Key points:
Build a “trust ladder”. The more you can constrain the agent’s output, the more you can trust it.
Instead of a mere “software factory”, try to create a “Michelin restaurant”. Create processes and tools that make beautiful code automatically.
Think about code reviews like sampling. You can’t afford to taste every dish, but you should spot check.
Ask “where am I the bottleneck? How can I get the agent to answer its own questions?”
Create inner loops and outer loops. Inner loops accomplish tasks. Outer loops coordinate inner loops and analyze the big picture.
Your skills should get smaller over time as agents improve.. Focus on workflows, not implementation details.
Connectors trigger new work. Connect your bots to Jira, GitHub issues, etc so they begin work automatically.
Make it easy for agents to verify their work. Create a CLI for your app and a skill that calls the CLI. Then agents can verify their work consistently and efficiently.
A good code base is easy to change.
Probably the best interview of 2026. @poteto and Matt are both awesome 😎 👏
55
68
791
56,537
elias tazartes 🥕🧑🌾 retweeted
My job no longer involves writing code.
My job:
- Choose and prioritize tasks
- Decide how to prompt
- Use instructions, skills, and tools to guide the AI
- Create verifiers so the AI can validate its work.
- Validate the results and iterate
My job:
- Choose and prioritize tasks
- Decide how to prompt
- Use instructions, skills, and tools to guide the AI
- Create verifiers so the AI can validate its work.
- Validate the results and iterate
67
63
1,031
41,009
elias tazartes 🥕🧑🌾 retweeted
I tracked down the exact bug that drained Payy. Easy to miss in the noise, but particularly scary because of how it happened.
TLDR; the code was correct, yet was drained anyway. Details below...
TLDR; the code was correct, yet was drained anyway. Details below...
Bitget hacked for $178M
Payy hacked for $1.8M
onchain is back baby
9
16
224
55,935
elias tazartes 🥕🧑🌾 retweeted
Just dropped a new DJ set, 30 mins of pure delight to blast in your living room
soundcloud.com/user2211512/bo…
soundcloud.com/user2211512/bo…
4
1
12
668