@moebiousi
iAccount based inColombia
About this account
- Account based in
- Colombia
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
Cineasta de mentiras · Escritor frustrado · Software engineer
Joined November 2008
- Tweets33.3K
- Following7.5K
- Followers858
- Likes112K
Pinned Tweet
"La libertad es como la felicidad: nunca se llega. Nunca se tiene completa."
—Pedro J. Gutiérrez (Animal tropical)
Kevin Vicent retweeted
Here's a 45-second TL;DR on Jev.
I find the core idea beautifully simple, but the video made it really hard to understand.
Hope you find it helpful.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Kevin Vicent retweeted
We tested @typesafeai on legal corpus.
In short: It works! It's worthy
Kevin Vicent retweeted
Doc-OCR router using Jev @typesafeai
A Jev-powered router that looks at a PDF page by page, decides which pages actually need OCR, extracts the rest locally.
Result: save cost on # OCR pages + speed
Kevin Vicent retweeted
what is Jev and how to use it, a deep dive ⤵︎
flaviocopes.com/jev/
Lo más difícil de desarrollar software con IA para negocios reales, no es pedirle que cree una API, diseñe o haga frontend + backend + base de datos. Eso la IA ya lo puede generar.
Lo difícil es entender al cliente: no siempre sabe lo que quiere exactamente, su problema suele ser más complejo de lo que cree y, si por él fuera, te pediría una interfaz con 50 botones y una tabla enorme al estilo Excel, sin mencionar que no tiene ninguna consideracion de como escalar un sistema real.
Ahí está el trabajo real del desarrollo actual: ser el intermediario entre la IA (la que genera el código) y el cliente (el que explica el problema).
Kevin Vicent retweeted
Jev by @typesafeai is free on Vercel AI Gateway until Sept 25.
Build with the fastest adopted model on the Gateway at no cost.
vercel.com/ai-gateway/models…
nitter.cf/vercel/status/21010773…
Jev was adopted faster than any other model in AI Gateway history.
In the first day, @typesafeai reached ~13% of teams, 2x the GPT-5.6 family and 6x Fable 5.1.
jev class of models have no generative decoder loop, have bounded outputs, utilize parallel inference, & have extreme compression/efficiency incentives. this makes them perfect to run on device (this will happen soon).
at that point you can put judgment into basically every existing interaction like notifications, email, camera, keyboard, home screen, sensors, shortcuts, accessibility, etc. with zero marginal cost & it would be near instant (users would never feel the lag of an llm).
more specifically you can frame lots of problems that are potentially too expensive for an llm into decisions, stuff like:
should i surface this?
is this unusual?
does the user care?
which action is appropriate?
is this interruption worth it?
did their intent change?
should this ui adapt?
this basically adds a new computational programming primitive for ~free.
"In the last 18 months, computer-use agents crossed from demo to deployable."
A year ago, the best computer-use model scored 42% on the standard benchmark for agents operating a real desktop.
Today's best: 85%. Human testers score ~72% on the same tasks.
The data on computer-use agents, from
@fabrisera2000, @seema_amble, and @zephratic: a16z.news/p/can-agents-use-a…
Kevin Vicent retweeted
Breaking: Browser Use + Jev = Ultrafast ⚡
Findings flights took 7s and cost only $0.0039 🤯
> new action space every step
> DOM state space
> small LLM fallback to type
(this video is at 1x speed btw)
Built a tiny open source browser agent. try it below ↓
Kevin Vicent retweeted
here’s everything I’m using in AI right now
Fable 5.1 (planning)
GPT-5.6 Sol (everyday projects)
GPT-6 Astra (hard PRs + plan audits)
GPT-5.6 Sol & Opus 5 (making prompts)
Jev (high-volume decision making)
Fable 5.1 (orchestrating)
Codex (vibecoding w/ computer use)
Grok 4.6 (vibecoding for fun/crons)
Cursor (projects + cloud agents)
Devin AI (cost-efficient hybrid coding)
GPT-6 Astra (building games + worlds)
Opus & Fable 5.1 (frontend/UI polish)
Figma & Paper (designing with agents)
Figma Motion (motion design)
Rive (interactive motion design)
Gemini 3.8 Flash (vision)
Supabase (backend)
Vercel (hosting)
ChatGPT Images 2.5 (image gen/editing)
Grok Imagine Image 2.0 (storyboarding)
Midjourney (viral image styles)
Gemini Omni 1.1 Flash (10s video gen)
MiniMax H3 (15s video gen)
Seedance 2.5 (30s video gen)
Meshy (image/text → 3D)
HeyGen (AI avatars)
Suno v6 (music gen)
ElevenLabs v3 (voices + sounds + SFX)
Grok Bot (clipping + captions + editing)
CapCut (quick video editing)
DaVinci Resolve (pro video editing)
Grok Bot & Hermes (autonomous agents)
Wispr Flow (voice to text)
Granola (meeting notes)
Obsidian (AI knowledge base)
Notion (notes for agents)
Grok (research + fact checking)
anything missing here? 🤣
i’ll give it a try and report back
We got computer use at home.
GLM drives this so well, I’ve been using it day and night and having it manage my computer and invoices.
Free and open source
Kevin Vicent retweeted
My Stanford course 𝗧𝗵𝗲 𝗠𝗼𝗱𝗲𝗿𝗻 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 starts next Tuesday.
This Github repository will hold all the assignments (and has the ones from last year).
Bookmark it. See you next week.
github.com/mihail911/modern-…
Kevin Vicent retweeted
Forward Deployed Engineer a pie de calle =>
javisantana.com/fde/
He intentado resumir 8 años de aprendizajes intentando no ser demasiado sofisticado ni añadir demasiado fuego artifial.
Está muy orientada a empresas de producto B2B, aunque creo que hay algunas lecciones que sirven en otros contextos.
Gracias a todo el mundo que lo ha revisado
Kevin Vicent retweeted
Top 15 AI Engineer projects for the 2026 hiring season.
If you can build these.
You're hired.
Project 1: Terminal Coding Agent
CLI agent that reads files, writes code, runs tests, self-corrects on failure.
→ Shows: You can build agents that operate real systems safely
Project 2: Custom MCP Server Suite
Build 3 MCP servers (database, filesystem, external API) with auth and rate limiting.
→ Shows: You understand the protocol connecting agents to tools
Project 3: Computer Use Agent
Browser automation that fills forms, extracts data, hands off to humans on CAPTCHA.
→ Shows: You can orchestrate agents on real-world interfaces
Project 4: Real-Time Voice Agent
Streaming voice assistant with sub-1s latency, interruption handling, tool calling.
→ Shows: You can build low-latency multimodal systems
Project 5: Multimodal Document Agent
Ingests PDFs with tables, charts, images. Reasons across modalities, cites sources.
→ Shows: You can handle enterprise data beyond plain text
Project 6: Agent Memory System
Short-term buffers + long-term vector recall + context compression across sessions.
→ Shows: You can make agents feel intelligent not amnesiac
Project 7: Eval-Gated CI/CD Pipeline
Automated evals that block deploys when quality drops below threshold.
→ Shows: You treat AI quality like code quality
Project 8: Agent Red Teaming Scanner
Automated prompt injection + jailbreak fuzzing against your own agents with fix reports.
→ Shows: You build safe systems before attackers find the holes
Project 9: Cost-Aware Model Router
Routes queries by complexity across model tiers, with caching and budget caps.
→ Shows: You understand unit economics not just capabilities
Project 10: Open Model Fine-Tuning Pipeline
Synthetic data generation + QLoRA fine-tuning + before/after evals on a domain task.
→ Shows: You can customize open models instead of renting closed ones
Project 11: Self-Correcting Agentic RAG
Query routing, hybrid search, reranking, critique loop that retries on low confidence.
→ Shows: You can build retrieval that fixes itself
Project 12: AI Observability Dashboard
Traces, token cost, latency percentiles, hallucination rate in one Grafana view.
→ Shows: You can debug production AI in minutes
Project 13: Multi-Agent Swarm with Consensus
3+ parallel agents vote and merge outputs, with conflict resolution and supervisor fallback.
→ Shows: You can orchestrate complexity without chaos
Project 14: On-Device AI Application
Quantized local model running in browser or edge, with offline fallback and sync.
→ Shows: You can ship private, zero-cost inference
Project 15: Agent-to-Agent Commerce Prototype
Two agents negotiate, transact & log an escrow-style audit trail.
→ Shows: You are building for the machine economy, not just demos
Most people stay stuck watching tutorials.
Builders get hired.
Bookmark & Repost!
Kevin Vicent retweeted
32% of companies just skipped a software purchase.
Not because the software was bad. Because their team built the equivalent in a week using a coding agent.
McKinsey's 2026 State of AI survey: 32% of organizations decided against buying at least one software product or feature because agentic coding tools made internal build faster than vendor procurement. In tech, that number is 41%.
The broader numbers: enterprises scaling AI agents jumped from 27% to 40% in a year. A shift that fast is structural, not seasonal.
What I keep noticing about this pattern.
The build-vs-buy decision used to be a capital question. Building requires engineers, time, and maintenance overhead. Most SaaS vendors win that comparison by default. An AI coding agent changes the math: a senior engineer plus Claude Code or Codex can now ship an internal tool in days that previously required a six-month procurement, an enterprise contract, and a year of onboarding.
The category where this hits hardest: internal tooling, dashboards, data connectors, light CRM features, workflow automation. The tools companies used to buy to solve specific operational problems.
There's a catch, though. McKinsey found only 37% of organizations report AI contributed to EBIT, unchanged from a year ago. Adoption is climbing fast. Financial return is not following as fast. And around 20% of organizations say token costs are constraining their use.
The bottleneck is shifting from model quality to run cost and maintenance. The orgs that figure out cheap, reliable agentic pipelines will have a real cost advantage over orgs still paying SaaS per-seat pricing for the same outcome.
nitter.cf/EvanKirstel/status/209…
32% of orgs have now skipped buying a software product because they built it internally with agentic coding tools. Enterprises scaling agents jumped from 27% to 40% in a year. Build vs buy flipped for the first time in two decades. Every SaaS vendor should read this. @McKinsey mckinsey.com/featured-insigh…
Kevin Vicent retweeted
Here's why we built and open-sourced Agentic Inbox: an email inbox you can host yourself with a built-in AI agent, running entirely on Cloudflare Workers 👇
Today, we’re announcing Ternary Bonsai 2 27B.
Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.
Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use.
Ternary Bonsai 2 27B is available today under Apache 2.0.