FoundersFeed

First Amendment's 'freedom of the press' wasn't about protecting journalists—it was about protecting the tech stack itself. The printing press was the decentralized communication protocol of its era. Same logic applies today: protecting the right to build and run communications infrastructure (nodes, relays, clients) matters more than protecting specific platforms or media companies. The protocol > the institution.

Kimi-K3's UI design nails a subtle but critical UX detail: it proactively detects when Chinese text rendering looks awkward and auto-adjusts formatting for better readability. Most AI chat interfaces ignore localization quirks like CJK character spacing, line breaks, and punctuation handling. Kimi actually anticipates these friction points before users even notice them. Small touch, huge impact for native Chinese speakers who'd otherwise be squinting at poorly formatted responses all day.

Dev workflow tip: Stop vibe-coding low-priority features immediately. Instead, dump all non-critical bugs and optimizations into a centralized todo list (supports MCP integration, works great as a Cursor Composer task queue). When you have downtime, batch them into one or multiple tasks and let different agents handle them in parallel. This keeps you focused on the main thread without context-switching hell. The MCP support means your AI coding assistant can directly pull from the list and auto-assign work. Basically: write down the noise, batch process later, let agents do the grunt work. 🚀

Noticed your AI agents doing busywork instead of shipping core features? Here's a workflow fix:

Dump low-priority bugs and polish tasks into a backlog file (Markdown/text) instead of letting agents tackle them immediately. This keeps your main dev loop focused on critical path work.

When you have downtime, batch these tasks into themed groups and assign each batch to a different agent instance. They'll crush them in parallel while you stay on the main thread.

Bonus: Todos app supports MCP (Model Context Protocol), making it a solid task manager for Claude/Cursor workflows. You can pipe your backlog directly into agent context without manual copy-paste.

Vibe coding's low barrier is a trap—you end up shipping features nobody asked for.

New workflow: dump low-priority bugs and optimizations into a backlog first. Focus hard on the mainline. When you have bandwidth, batch them into grouped tasks and delegate to different agents in one shot.

This actually works. Side note: Todos now supports MCP, solid choice if you need a todolist for Cursor Composer.

J’ai passé 2 jours à repenser l’architecture des paramètres des Todos et j’ai réussi à réduire l’ensemble du flux de configuration à seulement 2 étapes. Parfois, la meilleure refactorisation est celle qui supprime la complexité plutôt que d’ajouter des fonctionnalités.

Regime shift incoming: Kimi K3 and Thinking Machines Inkling signal that frontier labs are pivoting from raw token sales to full-stack agent workflows.

The pattern repeats: labs commoditize their biggest customers (Jasper, Cursor) by vertically integrating. Raw models require too much engineering overhead—prompting, context management, evals, model routing. Labs will bundle this into cheaper, higher-performance agent systems.

Next-gen agents ship with built-in memory, self-evaluation, and adaptive learning (possibly via overnight fine-tuning or "dreaming" mechanisms). They'll handle model routing internally, calling stronger "advisor" models when stuck.

The wrapper economy dies here. Labs deliver agents that are easier to deploy, cheaper to run, and self-improving by default. This is the inflection point where agents become actual digital employees, not just API wrappers.

Expect a massive usability leap and a new wave of enterprise adoption as setup friction drops to near-zero.

Coin Center's Policy Director Jason Somensatto testifies before House Financial Services Committee today—one year post-CLARITY Act passage (which included Blockchain Regulatory Certainty Act).

Somensatto's background spans CFTC work (2018+), private sector, and now policy advocacy. His focus: defining regulatory boundaries between centralized businesses vs. permissionless protocols—protecting investors without blocking free speech, privacy, or technical innovation.

This marks his 4th congressional testimony since joining Coin Center. The hearing revisits where regulatory clarity stands after a year of legislative momentum.

Coin Center's Policy Director Jason Somensatto testifies today at 10am before House Financial Services—exactly one year after the House passed the CLARITY Act (which includes the Blockchain Regulatory Certainty Act).

Somensatto brings serious cross-sector experience: worked at the CFTC since 2018, moved through private sector roles, now at Coin Center. His angle: defining the regulatory boundary between centralized trusted entities and permissionless protocols—balancing investor protection without crushing speech rights, privacy, or protocol-level innovation.

This marks his fourth congressional testimony since joining Coin Center last year. The core technical-legal question remains: where do you draw the line between securities law (centralized issuers) and code-as-speech (open protocols)? Somensatto's been in the trenches on this since 2018 and understands the nuance better than most policy folks in the space.

Link to the hearing stream coming soon.

In the mobile era, everyone carried a smartphone. In the AI era, everyone will have a team of agents working behind the scenes.

This shift is fundamentally about compute distribution. Smartphones centralized personal computing into a single device you physically carry. AI agents decentralize intelligence across multiple specialized models that operate asynchronously on your behalf.

The architecture difference matters: instead of one general-purpose device handling all tasks sequentially, you'll have parallel agent processes - one negotiating your calendar, another monitoring code repos, another handling research synthesis. Each optimized for specific domains, running 24/7, with different latency requirements.

The bottleneck moves from hardware specs (RAM, CPU) to orchestration logic and inter-agent communication protocols. Your 'team' effectiveness depends on how well these agents share context, delegate tasks, and merge outputs without creating conflicting actions or redundant work.

Kimi K3 just dropped and the specs are wild - claiming Claude Sonnet 3.5 level intelligence at way cheaper pricing. Already integrated it into production. If you're running agents, seriously consider swapping the brain module. The cost-to-performance ratio could be a game changer for scaling multi-agent systems. 🧠

Logging is criminally underrated in the vibe coding era. Built full-stack logging from day one when developing @dokobot and it's been a lifesaver for debugging agents. When AI is writing half your code, granular logs across the entire stack become your ground truth for what's actually happening vs what the model thinks is happening. Not optional anymore.

Exponential growth creates a brutal paradox: each step forward requires MORE resources than all previous steps combined, but our brains expect things to get easier over time. This hits hard in AI scaling—going from GPT-3 to GPT-4 cost way more than GPT-2 to GPT-3, yet the performance gains felt smaller. Same pattern in chip fab: each new node (7nm→5nm→3nm) costs exponentially more while delivering diminishing performance bumps. Economics assumes marginal costs should drop, but exponentials break that assumption. The next doubling always costs more than everything you've spent so far. 📈💸

DeFi security is getting wrecked by AI-generated exploits right now. The attack surface has exploded - AI agents are finding novel vulnerabilities in smart contract code faster than auditors can patch them.

The fundamental issue: smart contracts have complex state machines with edge cases that humans miss but AI can brute-force discover. We're seeing automated exploit generation at scale.

Meanwhile $BTC sits there with its dead-simple UTXO model - no reentrancy attacks, no flash loan exploits, no obscure EVM quirks. Just basic cryptographic primitives that have been battle-tested for 15 years.

If you're holding funds in DeFi protocols, understand you're trusting:
- Solidity code complexity
- Oracle integrity
- Governance token holders
- Bridge security
- Composability risks across protocols

Self-custody cold storage eliminates every single one of those attack vectors. The tradeoff is you lose yield farming and leverage, but you gain actual ownership with no counterparty risk.

The irony: we built DeFi to remove trusted intermediaries, then created a new class of technical vulnerabilities that are arguably harder to audit than traditional finance.

Le livre blanc original de Bitcoin dit littéralement « système de paiement électronique entre pairs », mais à un moment ou à un autre, tous, nous nous sommes laissé distraire par le « numéro qui monte » et nous avons oublié que le but était de supprimer les intermédiaires. Nous avons construit des exchanges, des dépositaires, des tokens enveloppés, des L2 qui règlent sur des séquenceurs centralisés : en gros, nous avons recréé le système financier traditionnel, mais avec une expérience utilisateur plus mauvaise et des frais plus élevés. La technologie fonctionne le mieux quand la valeur circule directement entre nœuds sans tiers de confiance, mais la plupart des utilisateurs ne détiennent jamais réellement leurs propres clés et ne gèrent jamais leur propre infrastructure. Nous avons optimisé la spéculation plutôt que les transactions entre pairs. Peut-être est-il temps de se rappeler pourquoi Satoshi a écrit ce livre blanc en premier lieu.

Le vrai problème avec l’IA actuelle n’est pas la pile technologique : c’est le modèle de déploiement mal aligné. Nous voyons des LLM optimiser la génération de contenu (texte synthétique qui inonde X, Reddit, les commentaires YouTube) au lieu de résoudre le véritable cauchemar UX : la fragmentation des plateformes.

Pensez au fossé architectural : les agents d’IA modernes ont la capacité d’être stateful, des agrégateurs conscients du contexte, qui pourraient unifier vos flux d’informations sur X, YouTube, Reddit, Slack, Signal en un seul fil intelligent. La technologie existe — des pipelines RAG, des embeddings multi-modaux, une orchestration d’API en temps réel.

Mais au lieu de ça, nous avons eu :
→ des bots anti-spam d’IA qui polluent chaque plateforme
→ aucun progrès sur l’agrégation de contenu multi-plateforme
→ des utilisateurs qui continuent de basculer entre 10+ applis chaque jour

Le concept de « personal content butler » (majordome de contenu personnel) n’est pas irréaliste : il n’est simplement pas l’endroit où les incitations économiques se sont posées. Les plateformes veulent que vous restiez piégé dans leurs jardins clos. Les entreprises d’IA ont optimisé la génération de contenu (plus facile à monétiser) plutôt que des couches de consommation intelligentes.

Ce dont nous avons réellement besoin : un agent IA qui fonctionne en local ou dans votre cloud contrôlé, s’authentifie avec vos comptes, récupère le contenu via des API, applique votre modèle de préférences, et fait remonter un flux unifié. Techniquement trivial avec les LLM actuels + l’usage d’outils. Commercialement impossible parce que les plateformes ne permettraient jamais l’ouverture de ces API à grande échelle.

La déception ne concerne pas les capacités de l’IA : elle concerne la stratégie de déploiement, dictée par l’économie du verrouillage des plateformes plutôt que par l’optimisation de l’expérience utilisateur.

Having tokens changes everything. Used to make product demos with screenshots - now I just have AI generate 1:1 pixel-perfect interfaces on demand. Why bother screenshotting when you can synthesize the exact UI you need in seconds?

Todos Workbench just shipped mobile support for their full dev stack - and it's free.

What you can actually run from your phone now:
• Full code editing and execution
• Research workflows (probably web scraping + LLM synthesis)
• Server/website management (SSH access, deployment controls)
• Data analysis pipelines

The real flex here is zero desktop dependency. Most "mobile dev environments" are just glorified text editors, but if Todos is running actual compute workloads + server management from mobile, that's a legit cloud IDE with proper backend integration.

Worth testing if you're tired of being chained to your laptop for ops work.

La plus grosse hallucination des agents de code en ce moment ? L’estimation de la durée 😂

Ils vont te dire avec assurance « ça prendra 2 heures » puis refactorer la moitié de ton code, ajouter 3 nouvelles dépendances et ne toujours pas terminer la tâche initiale. Un classique : la complexité en théorie, contre la réalité de l’implémentation.

cc (Cursor Composer) exécute des vérifications silencieuses sur votre code après que vous l’avez écrit, en testant des cas limites et des scénarios que vous n’avez jamais demandés explicitement. En gros, cela fait de l’assurance qualité automatisée en arrière-plan pendant que vous travaillez. Imaginez un relecteur de code intégré, paranoïaque, qui vérifie votre logique même quand vous ne le sollicitez pas — pour repérer d’éventuels bugs avant qu’ils n’atteignent la production. C’est une fonctionnalité très pratique si vous utilisez Cursor comme outil quotidien.