Mauro Toscano 🟩 (@mauro_aligned) profile picture

CTO of @alignedlayer Opinions are from my AI

Joined December 2021
Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)Photo by Mauro Toscano 🟩 (@mauro_aligned)
We have so many things to fix and improve in the world that even with AI it’s hard
we are in the non stop exponential age. everything is chaos and opportunities. but somehow people still do lines at banks to even ask how much money they have in their account. the only thing i'm sure is that we will live in a pretty absurd world.
1
2
248
And run your reviewers in agentic mode, inside Codex or Claude Code. It increase the cost but you can hide security vulns if you only do a diff analysis or can hide stuff from the agent
Right now, this is my impression of models as reviewers. We’re still collecting data. Opus xhigh: Excellent. Finds even subtle issues in our more complex code (still not Fable level). Opus high: Very good, but may miss some particularly complex issues. GPT Sol: Very good and relatively cheap. Kimi: Very good, and sometimes finds issues the others miss. DeepSeek Pro: Good. GLM: Mid. It’s fine, but doesn’t seem to provide any unique insights, and it’s as expensive as more frontier-level models. DeepSeek Flash: Incredible for reviewing everything that isn’t security-related, and extremely cheap (haven’t tried using it to find security issues yet). Grok: Awful. Burns a lot of money and doesn’t find anything particularly interesting. Might as well use Fable. Gemini: Bad. Burns a lot of money and barely finds anything the others miss. A bit better than Grok, though. Then If you want a judge to remove false positives. As judges go, opus high seems the best of the ones I tried but gpt does the job decently for half the price.
3
2
136
Right now, this is my impression of models as reviewers. We’re still collecting data. Opus xhigh: Excellent. Finds even subtle issues in our more complex code (still not Fable level). Opus high: Very good, but may miss some particularly complex issues. GPT Sol: Very good and relatively cheap. Kimi: Very good, and sometimes finds issues the others miss. DeepSeek Pro: Good. GLM: Mid. It’s fine, but doesn’t seem to provide any unique insights, and it’s as expensive as more frontier-level models. DeepSeek Flash: Incredible for reviewing everything that isn’t security-related, and extremely cheap (haven’t tried using it to find security issues yet). Grok: Awful. Burns a lot of money and doesn’t find anything particularly interesting. Might as well use Fable. Gemini: Bad. Burns a lot of money and barely finds anything the others miss. A bit better than Grok, though. Then If you want a judge to remove false positives. As judges go, opus high seems the best of the ones I tried but gpt does the job decently for half the price.
1
1
1
8
512
We have been doing our own work with this and back tracking if other AIs, prompts or effort levels would give us better results. I will post a small tier list of configs though, that we are still experimenting with
Backtested 30+ AI code reviewers on 10 merged PRs Lesson: mix vendors, don't buy the biggest model. Claude Code+Haiku with codex review+GPT-6.2 Sol: 19/26 important bugs, $0.42/PR Opus 5.5 medium alone: 12/26, $1.57 Medium effort beat high/xhigh in most cases
1
4
388
I want my agent to pay for stuff, I’m tired of it suggesting “make an account, put your credit card, pay for a subscription, then cancel”
We're excited to announce that 4Mica v2 is now live on @base! This upgrade sets the stage for dynamic, demand-based pricing and seamless discoverability for API sellers, giving agents access to a larger library of tools. Building tooling for AI agents? Reach out to chat about how 4Mica can help. Blog: 4mica.io/blog/mainnet-v2-cyc…
2
9
349
This is false. OpenVM, LambdaVM, and Airbender all work without relying on elliptic curves or lattices.
Good quote towards the end of this clip by @srinathtv: "[There's really only one hash based zkVM]... leanVM from the Ethereum Foundation" the rest of the "hash based zkVMs" are really "hash and more" zkVMs. well put.
1
2
1
20
2,424
This question has been at the center of our work over the last few months with @3miLabs, @class_lambda and @alignedlayer We hope it helps the community understand what "formally verified" really means, and decide whether humans should still be able to read what's being proven
Should humans be able to understand a zkVM's formal spec, or is "Lean checked it" enough? To show what formal verification of a VM actually involves, we walked through a toy one, step by step
3
5
262
Should humans be able to understand a zkVM's formal spec, or is "Lean checked it" enough? To show what formal verification of a VM actually involves, we walked through a toy one, step by step
1
1
13
561
These weeks we have been creating different versions of the LambdaVM that were never materialized. It's amazing to run all the "what if" scenarios We are finding some optimal configurations that we missed by doing local optimizations of some steps of the pipeline
5
249
Amazing work!
Happy to announce my latest work with @skominers and @SuccinctJT ! eprint.iacr.org/2026/2056 TLDR: we transfer the techniques from better.codes to yield 5-10% proof size reductions in several SNARKs today, and supply an efficient decoder for knowledge soundness 🧵/ n
176
We need more instances of just putting a lot of people together to solve stuff
2
8
623
The quantum threat is becoming more real. The biggest tech companies are working to migrate to PQ. It’s going to be a fun ride
2
174
Mauro Toscano 🟩 retweeted
I want to congratulate @WillyDuran65 on his new term as dean of the Faculty of Exact and Natural Sciences of the University of Buenos Aires. Willy thanked Francisco Strambini and me, and he also thanked @class_lambda for its contributions and its relationship with the university. I value that a lot. Most people on X have a somewhat wrong image of what Lambda is. Many believe we’re a devshop that works in crypto, when we have dozens of successful products in different industries and most of our core members are mathematicians, physicists, mechanical engineers, chemical engineers, civil engineers and industrial engineers and yes obviously computer scientists and computer engineers. Lambda is and has always been deep tech. If you want to build SaaS (something that we also do), you don’t need to understand what a Koopman operator is, finite element method or what Riemann zeta function is. But if you want to build companies at the frontier of science and engineering, you need people who understand things deeply and can work on hard problems. That’s why Lambda’s relationship with the University of Buenos Aires, with the Faculty of Engineering and particularly with the Faculty of Exact and Natural Sciences is so important and strategic. For me, the relationship with the university has always been strange and very personal. I’m a college dropout who, more than 15 years later, came back to reform the same degree program I had dropped out of. It’s a strange circle. I left the university without a degree, but part of the university is still shaping the way I think. I spent more than 20 years in the private sector learning how to build companies, products and technology. Now I’m using everything I learned to build a new degree program at the University of Buenos Aires. I want to create an elite of builders who want to innovate. Not an elite based on money, family or status. An elite based on knowledge, work and the ability to turn science and engineering into companies and good public policies. For me, this is not about proving that I never needed the university. It’s the opposite, it’s about giving back, closing a circle and helping create the education that I believe the next generation of builders needs. If we want to create a better world, we need to create companies that are at the frontier of science and engineering and that work with universities and the public sector. We also need to destroy this idea, that part of the left has, that companies go to public universities only to steal their knowledge or to destroy them. Real deep tech companies need strong universities. A deep-tech company has nothing to gain from a weak university. But this works both ways. Universities also need a big and capable private sector with the ability to innovate, turn knowledge into products and give back to the universities that made that innovation possible. Universities have the right environment to do research that is long term and doesn’t depend on the results of the next quarter. This is something that part of the right in Argentina doesn't understand. Companies need the public universtiy that is the one creating the most amount of professionals for our country and doing the best international research. The University needs money to be able to give freedom to researchers to work on difficult problems. Then companies have the ability to turn that research and knowledge into products, jobs and new industries. They have different roles, but they need each other. This shouldn't be an ideological debate but a practical one. Willy and I have our differences, as all good friend have, but we also have a very strong bond. We believe in science. We believe the state should invest in it. We believe this is better for the country, for universities and for companies. It’s a win-win. That’s why I value my relationship with Willy so highly. Even with our differences, we agree on what we want. Thanks Willy and I hope you keep doing an amazing job.
2
5
3
28
3,501
Market efficiency goes up
growth of 20% in a week. 1.13 billions went through of our propamm infra the last week. up only.
1
6
475
Mauro Toscano 🟩 retweeted
The state of Nuevo León, the richest in Mexico, is using @ethrex_client and the @alignedlayer white label solution wallet. Every important government procedure is being built on top of our stack, and millions of procedures every year will be powered by it, all secured with ZK and Ethereum. Congrats @sovraio and @class_lambda. Technology developed by the Ethereum ecosystem becoming infrastructure that millions of people can rely on in their everyday lives. Most citizens may never need to know what runs underneath, but they will benefit from systems that are more secure, transparent, efficient, and verifiable. For the last two years, I’ve been saying that we’re developing multiple solutions that use the technologies Ethereum pioneered to make states more powerful and citizens more free. We're working with multiple strategic partners around the world. Ethereum is not only building a new financial system, it is giving us the tools to build better institutions and better public infrastructure. This is only the beginning. We’ll be sharing much more in the months to come. And a special congrats to @fernandezdiego, our CEO in Sovra, for all the hard work that helped make this possible. We are closing more and more big deals, Diego is a deal making machine. I've learnt so much from you, I'm proud of marching under your command.
3
7
2
31
4,021
Mauro Toscano 🟩 retweeted
four months ago propamms was just an idea we had. today $300M a day flows through our systems in ethereum and are powering thousands of trades per day. now, I'm starting a new company and product again, and I'm just as sure about this one. urgently hiring, part time or full time in Buenos Aires (I might accept people from other parts of the world): 1. RF/microwave engineer 2. Radar DSP engineer 3. GNC (guidance, navigation & control) engineer if you've worked at INVAP, CONAE, CNEA or Instituto Balseiro, in automotive radar or at TI/NXP/Infineon, or you're part of a university radar lab or the microwave/ham radio community, or you know someone who is, DM me and email [email protected] mentioning this tweet. i’m running this project and I'll be on it 24/7. open to giving equity in the new company if the MVP goes well. unless health or family problems appear, i’m gonna do everything i can for this to be a global success too. i believe it can be great for society too. we already have a garage as a new office for this project. RT and share if you can, people looking for a new job or project might find it that way. Thanks!
20
21
5
120
23,520
Congrats to 3MI, Amit, Sundas and Tomer!
tl;dr: we present an interpolation attack on 28 out of 31 rounds of Poseidon in practical time via a novel technique for partial round skipping; see ia.cr/2026/1692. 🧵⬇
2
7
588
Mauro Toscano 🟩 retweeted
Over the past months, Sovra, @class_lambda and @alignedlayer have been working together on Propia, a public protocol for verifiable digital credentials. Propia starts from a simple premise: trust established by one institution should be reusable by others, without requiring shared databases or dependence on a private intermediary. This has become increasingly relevant as AI agents begin to participate in everyday operations. They submit requests, exchange information, authorize transactions and interact with other agents. These interactions need a reliable way to establish who an agent represents, what it is authorized to do and whether the claims it presents remain valid. Rebuilding that verification process within every organization is costly and creates isolated systems. A shared protocol allows credentials to be issued once and independently verified wherever they are needed, including by machines acting on someone’s behalf. For institutions, this means fewer repeated checks, fewer custom integrations and lower operating costs. As more institutions participate, more credentials can be reused and the protocol becomes more useful for everyone. We believe this infrastructure should be open and collectively governed. Today, we are publishing the whitepaper as a public proposal and inviting others to discuss it, improve it and build on it.
4
4
7
14
18,480