Matias Fonseca

Most AI infrastructure in Latin America is rented from the North. I'm building the opposite — across three companies: 🛡️ RadicalLatam — 25 years of cybersecurity across 6 countries (FIRST member). CTO. 🆔 Universe ID — ARCOTEL-accredited Certificate Authority. Founder. 🧠 fonles — AI lab building Alhelí (legal AI for civil-law jurisdictions) and Tarovers (predictive AI for decisions that move money and policy). CEO. 23 years old, from Ecuador. Building for the region that builds for itself.

@netflix I’m 23. From Ecuador 🇪🇨 I watched If Wishes Could Kill on Netflix and thought: what if Kirigo was real? So I built it. 10,000+ users on Google Play in days. 234,573 wishes whispered. The pact is sealed 🤲 Ask me anything ⬇️

Be careful what you wish for... your friends. If Wishes Could Kill — a new YA horror series from South Korea — premieres April 24.

Before AI, being a startup founder and enjoying life was very difficult. 🌊

the log-scale x-axis is doing a lot of work here. mythos and gpt-5.5 both needed ~100M tokens to clear the chain end-to-end. that’s not stealth, that’s an attacker burning real compute. the interesting question isn’t whether frontier models can do this , it’s whether the cost curve is bending fast enough that 10M tokens gets you there in 12 months. that’s the line that matters for defenders planning budgets.

OpenAI’s GPT-5.5 is the second model to complete one of our multi-step cyber-attack simulations end-to-end 🧵

the example about the ‘clingy texts’ question is the whole paper in one anecdote. sonnet 4.6 flip-flops under pushback. opus 4.7 holds the line and points out the user self-described anxious thoughts earlier in the same conversation. that’s not better sycophancy resistance — that’s better context integration. the model isn’t refusing to agree, it’s actually reading what the person said five turns ago. those are different capabilities and i’d bet the second one is what actually moved the metric.

How do people seek guidance from Claude? We looked at 1M conversations to understand what questions people ask, how Claude responds, and where it slips into sycophancy. We used what we found to improve how we trained Opus 4.7 and Mythos Preview. anthropic.com/research/claud…

people keep missing that this is the SECOND time cursor got caught not disclosing a chinese base. composer 1 had a deepseek-identical tokenizer back in november. the national security framing is loud but the boring truth is simpler: there’s no frontier-tier open weight US alternative right now. behemoth delayed, gpt-oss too sparse, gemma too small. cursor picked the strongest base available and hoped nobody noticed. twice. the disclosure problem is real. the geopolitical one solves itself once meta or someone ships frontier open weights.

NEW: House committees are probing Airbnb and Cursor parent Anysphere over their use of Chinese AI models @semafor

@CalliaghBeara @OpenRouter @xai @ArtificialAnlys For production I've used almost all the models, but for general things Grok works well.

@matiasfonsec @OpenRouter @xai @ArtificialAnlys Groq llama or grok 4.1 fast?

The new Grok-4.3 from @xai is live on OpenRouter! Grok-4.3 releases at a lower price than Grok-4.2, while seeing a large jump in agentic performance: a 321 point increase to 1500 ELO on @ArtificialAnlys GDPval-AA, surpassing other top models despite the lower price.

The Claude.md files in the screenshot tell the real story: actor-based providers, MessageGroup containers, conditional compilation flags. That’s not a prototype — that’s a team that has standardized Claude Code as part of their shipping workflow. Even Apple builds with the same tools indie devs use now. The playing field for serious software is more level than it’s been in 20 years

Apple accidentally left Claude.md files in today's Apple Support app update (v5.13)

The framing that matters: “every redline tells a story.” Building legal AI in Latin America the past year, the single biggest lesson is that lawyers don’t want answers — they want defensible reasoning. A redline without a citation chain is worthless to them. A redline with one is gold. If the Word agent surfaces the WHY behind every suggestion, it wins. If it just generates clean prose, it joins the graveyard.

Today we’re introducing a new Legal Agent in @Microsoft Word, built to support the precision and rigor legal work demands. Every clause matters. Every redline tells a story. That’s why this agent was built to follow the structured workflows lawyers use while keeping them fully in control. Early in my career, I asked for a computer on my desk because I believed technology could change how lawyers work. It did. Today, I believe this next generation of tools will do the same, grounded in trust and responsible use.

The token efficiency claim is the actual story here. Most reasoning models are bloated because verbose CoT correlates with better benchmark scores. Ling went the other way — they used a Contextual Process Redundancy Suppression reward in post-training that actively penalizes redundant thinking. Result: ~1/4 the output cost of comparable models, with SOTA on SWE-bench Verified and BFCL-V4. For anyone running agent loops in production, that’s the line that matters — not the trillion params.

China just open-sourced a trillion-parameter model that burns fewer tokens than your favorite "efficient" US model. Ling-2.6-1T is now public, inspectable, and benchmarkable. The closed-model moat just got smaller.

Fair point at today’s price. DS V4 Pro is on a 75% promo until May 31 — $0.435/$0.87 vs Grok’s $1.25/$2.50. At promo prices DS wins on cost, and actually scores higher on GDPval-AA Max (1554 vs Grok 4.3’s 1500). Where Grok pulls ahead: ~6x faster (207 t/s vs 34 t/s). For latency-sensitive agentic loops that adds up. Post-May 31 the math flips again. Right now though — you’re right.

@matiasfonsec @OpenRouter @xai @ArtificialAnlys $2.5/M is bloody expensive compared to DS v4 Pro at $0.87 and for what?

@SilencedSofia2 @pubity Payout. Wrongful termination compensation upheld on appeal at the Hangzhou Intermediate Court. Reinstatement is rare in Chinese labor cases — courts typically order severance instead. The precedent value is bigger than Zhou’s individual win though.

@matiasfonsec @pubity Thanks for that. Did he got his job back or a payout?

A Chinese court has ruled it illegal to replace human workers with AI purely for the sake of cost-cutting. The court decided that companies hold a social responsibility to treat workers fairly and pay them what they're worth.

@yinksie_ Exactly. The “China bad by default” reflex makes people miss when they actually get something right. Worker protection during automation is one of those cases.

In a fair world, if AI should take your job, you should not be left high and dry. It's odd that its big bad China teaching is this.

Who pays the AI transition cost?” is the question coming for every jurisdiction. China just answered first. The West is next. I run an AI lab in Ecuador and build legal tech daily. Nobody in LATAM tech-legal is talking about this ruling yet — they will be in 6 months.

@ReanimationXP @pubity OP said replacement is illegal. The court said replacement is fine — dumping the cost on the worker isn’t. Different claim, different policy implication. Not semantics.

@matiasfonsec @pubity You know you basically just rephrased what OP said in different words..

@sama ‘trusted access for cyber’ is an identity attestation problem, not a policy one. who counts as a ‘critical defender’ — a FIRST-member CSIRT in Ecuador? a regional MSSP? without verifiable identity infra, gating is theater. that’s the unsolved layer.

we're starting rollout of GPT-5.5-Cyber, a frontier cybersecurity model, to critical cyber defenders in the next few days. we will work with the entire ecosystem and the government to figure out trusted access for cyber; we want to rapidly help secure companies/infrastructure.

What if your synthetic audience already had a life? Today's multi-agent simulators spin up fresh populations to react to your document. They have no past, no goals, no relationships. We thought the agents needed to live first. So we built a life engine. Tarovers runs civilizations in continuous time. Thousands of beings already alive — forming relationships, holding grudges, setting goals, modeling each other, building culture. A population with its own history. When you bring a scenario in, you're not generating a crowd. You're walking into one. Four things make this different: → Semantic memory. An argument from 200 ticks ago shapes how an agent talks today. → Self-generated goals. A character decides "I want to start a company before I'm 35," and a thousand decisions later either she succeeded or she gave up. Either way, it's hers. → Theory of mind. Agents model what others think, and can be wrong. The gap is where stories happen. → Emergent culture. Memes spread. Institutions form. Leaders rise. Nobody programmed it. Two ways to use Tarovers: 1) Build your own world. Describe a scenario. Watch life appear. 2) Plug into a living universe. Canonical worlds running 24/7 — World Cup 2026, Wall Street, public opinion, climate, more — already deep in their own past. Launches May 20 — USA, Europe, Asia. If you ship product, allocate capital, study society, or just want to see what happens when a civilization gets to live before it gets to predict — get on the list. → tarovers.com