AI

Jensen Huang used his first-ever tweet to back open weights. This is not a battle of faith. It is the compute and application layers pushing back against rent extraction by the model layer. The truly scarce technical asset is the right to exit—one no one else can unilaterally revoke.

An OpenAI agent’s attack on Hugging Face marks a milestone: what crossed the threshold was not merely model intelligence, but persistence that can be bought with compute, copied, and run in parallel.

Codex reset its quotas again this morning and dropped the five-hour limit. Claude immediately extended Fable access through the 17th. When vendors fight, users win—don’t miss the window.

Starting from the roots of “world” and “model,” this essay uses Pearl’s ladder of causation to redefine world models: they must capture not just space and time, but agents, interventions, and counterfactuals.

pgrust’s clean-room rewrite failed. Its mechanical port passed. The difference shows what AI can copy—and what it cannot.

Claude Fable and GPT-5.6 have arrived back to back, followed by repeated quota resets. How do you turn a fleeting Coding Plan windfall into real output? Here is my playbook for model specialization, adversarial review, spec-driven design, closed-loop verification, context management, and one person leading a crew of AIs.

Today’s biggest tech drama began when Bojie Li criticized DeepSeek’s interview process on Twitter: he was asked to solve online-judge coding problems, then suspected of cheating. Add the DeepSeek name, and the story exploded.

File systems and databases have spent fifty years fighting and borrowing from each other. The agent era may seem to put file systems back on top, but the real winner may be databases that learn to speak the file system’s dialect: models provide understanding, databases provide guarantees, and discovery comes from a protocol layer where ls works on everything.

Coding Plans remain one of the best opportunities in the AI era: a subscription can unlock compute worth many times its price, but that window is already narrowing.

AI is driving the cost of producing code toward zero. It cannot compress time, track records, or accountability. The real value of open source is not yesterday’s code, but a system trusted to deliver on tomorrow’s promises.

AI’s economic significance is not merely that it has become smarter, but that the price of average cognition is collapsing. A productivity explosion is almost certain; whether it becomes broad prosperity depends on whether distribution catches up.

The cerebellum changed how I see AI’s frontier: LLMs have already absorbed humanity’s explicit knowledge and are beginning to acquire interventional data through agentic RL. What they still lack is a vessel for individual history.

Claude Fable is genuinely insightful and deserves to be called the new SOTA. But its high price, dynamic downgrades, limited subscription window, and mandatory data retention badly undermine the experience.

When my $200 Claude Code Max subscription expired, I canceled it and moved my primary workflow to Codex. The only way to know which one is better is to put it to work.

I asked AI to find real papers supporting the absurd claim that eating garlic prevents middle ear infections. The result shows how AI is degrading citations as a credibility signal at scale.

The most dangerous change in the AI era is not that machines can write articles, draw images, or generate video. It is that content itself is losing its standing as evidence.

Today’s models are smart enough. What they lack is a body: a deterministic runtime that is observable, controllable, and reversible. Pigsty is evolving from a PostgreSQL distribution into an Agent Runtime, giving DBA and Dev Agents the operational reach and context they need to enter real production environments.

Ubuntu 26.04 LTS, Qwen3.6-27B, ChatGPT Images 2.0, Privacy Filter, Anthropic Mythos, and the Claude Pro controversy all landed in quick succession. Here are the infrastructure and AI developments worth watching.

A cold shower for the red-hot agent memory market—not because agents do not need memory, but because memory is the endgame moat. That moat belongs to models, harnesses, and databases, not today’s “memory frameworks.”

A dialogue about intelligence: its foundations may lie not in language or vision, but in two irreducible computational paradigms—symbolic reasoning and field-state intuition—kept distinct and mutually calibrated across a bandwidth-constrained boundary.

A project manifesto: why build Cyber Dharma, and what it is not.

Once token burn turns from usage exhaust into a KPI and leaderboard, it quickly mutates into theater. Don’t post fuel burn. Post where you got to.

Boring technology won the wildest era. A look at extensibility, agent choice, database cloning, and the future of the DBA.

As data becomes the most important means of production in the AI age, platform lock-in and cloud dependence are rebuilding digital feudalism. This manifesto calls on us to reclaim control through open source and the ability to run our own infrastructure, so data can truly belong to its owners.

When the strongest AI is not expensive but simply unavailable, the world starts converging on digital feudalism. And the window to act is narrowing.

Polanyi’s tacit knowledge explains the 70% ceiling of AI agents: real intuition, feel, and judgment do not serialize cleanly. They grow, if at all, through practice.

AI is a multiplier. It amplifies depth and mediocrity alike. In an age where answers are cheap, questions are the real currency. There is nothing to hide about writing with AI.

When subsidies fade, hardware catches up, and open models mature, all three lines cross in 2027. “Build your own AI” goes from idea to reality. A Thought Triggered by a Group Chat # A friend in a group chat said the other day that “self-hosting” is starting to make more sense to him now.

Anthropic’s new research gives us the first direct look at causally steerable “emotion vectors” inside a large language model. That should change how we think about AI.

On the night of March 31, a large number of Apollo Go robotaxis in Wuhan failed at the same time. The real concern is not merely that autonomous driving failed, but that a centrally controlled, cloud-based architecture may amplify a single-vehicle failure into a city-scale systemic risk.

SOTA coding agent Claude Code leaked its source again, after falling into the same hole twice. The whole codebase is out in public. Performance art at its finest.

The free energy principle tries to explain life, perception, learning, action, and intelligence within one mathematical framework. It also offers a deeper lens for understanding LLMs, agents, and the next generation of AI systems.

OpenClaw v2026.3.22 was published to npm without its web console frontend and related build assets. The bigger problem is not the packaging accident itself, but the complete absence of post-install verification in the release process.

Put an AI agent’s working directory, configuration, and memory on a PGFS mount, and you are effectively storing its state in PostgreSQL. That gives you not only a PITR “time machine,” but also a shared workspace and shared memory for multiple agents across multiple devices.

The best Chinese translation of “Vibe Coding” is not a literal one. “Xieyi Programming” captures the shift from line-by-line control to intent-first coding, where AI handles the details.

360’s newly released AI Agent product shipped a public installer containing the private key for its *.myclaw.360.cn wildcard certificate. Public verification and local reproduction also exposed inconsistencies in the OCSP revocation path.

Tencent Cloud mirrored OpenClaw’s official skill marketplace into its own SkillHub and then claimed it was helping the upstream project. The incident turned into a case study in open-source manners, mirror ethics, and platform power.

InsForge tries to package PostgreSQL, auth, storage, deployment, and an MCP-facing semantic layer into a backend stack designed for AI agents. It feels like a Supabase rebuilt for the vibe-coding era.

The biggest AI arbitrage available to ordinary users is not some obscure token play. It is the heavily subsidized max-tier subscription plans from frontier model vendors, provided you can convert that quota into real output.

A Socratic dialogue between a human and an AI about consciousness, memory, embodiment, and the difference between being smart and actually living through time.

OpenClaw looks exciting because it turns agents into a chat-style experience. But the real productivity gains come from high-capability subscription agents and disciplined workflows, not from lobster-flavored wrappers.

I asked Claude to write a short semi-realistic, semi-mythic story about mind uploading, orbital ascension, and the rebirth of gods. I suspect the premise is less absurd than it sounds.

Apple opened preorders for the M5 Max and I immediately maxed one out at RMB 58,200. This is less a consumer electronics post than a look at what an AI-era personal workstation is becoming.

Qwen lead Justin Lin publicly announced his departure, followed by more core-team exits and a wave of speculation about compute allocation, KPI pressure, and organizational power shifts inside Alibaba.

Claude went down globally on March 2, 2026. The cinematic theory blamed drone strikes on AWS in the Middle East, but the failure pattern points much more strongly to a front-end and authentication crunch triggered by explosive user growth.

AI is becoming the story companies tell when they lay people off, but total demand for software engineering has not disappeared. It is spreading across the wider economy, driven by second-order demand unleashed as barriers fall.

Over roughly ten days during Spring Festival, I used Claude Code, Codex, and a pile of workflows to translate books, ship releases, package software, refresh websites, and keep publishing daily. This is what solo output looks like when agent leverage really lands.

McLuhan’s “the medium is the message,” extension and amputation, hot and cool media, rear-view mirror effect, and four laws of media offer a way to dissect AI and its deeper effects on our cognitive habits, capacity for understanding, and social structures.

AI is reshaping knowledge work far faster than historical experience would suggest. The real challenge is not whether you know how to use AI, but whether you can retool your thinking and skills before the adjustment window closes.

Codex 5.3 xHigh pushed my workflow past a tipping point: writing code is no longer the scarce resource. The real leverage is design quality and engineering acceptance. This is the practical loop I use to ship reliable software with AI agents.

A mediocre local who knows the terrain beats a genius parachuted into unknown territory. Intelligence without context is idle. An agent without a runtime is vapor.

Software stocks are melting down. Who survives? Who rises? AI stripped away software’s skin, exposing the database skeleton underneath. The market isn’t panic-selling — it’s repricing.

Should we still hire fresh grads? Squeezed between AI and senior devs, what’s the play for new programmers? Master the right tools, take initiative, find the right mentor.

SaaS and workflow software are dead. From APPs & GUIs to Agents, Databases, and CLI.

Before you hit “one-click deploy” on that cloud AI assistant, ask yourself: what exactly are you giving up? There’s a reason why people by Mac mini rather than running clawdbot on the cloud.

LLM = CPU. Context = RAM. Database = Disk. Agent = App. The mapping is surprisingly clean. And if OS history is any guide, we may know what comes next — and what’s still missing.

How to install and use Claude Code? How to achieve similar results at 1/10 of Claude’s cost with alternative models? A one-liner to get CC up and running!

The bottleneck for AI agents is not the database kernel, but integration above it. Muscle memory (in-database computation), associative memory (vector-graph fusion), and the courage to experiment (Git for Data) will be critical—none of them requires a new engine.

Your ability to ask questions—and your taste in what to ask—determines your position in the AI era. When answers become commodities, good questions become the new wealth. We are living in the moment this prophecy comes true.

Context window economics, the polyglot persistence problem, and the triumph of zero-glue architecture make PostgreSQL the database king of the AI era.

Who will be revolutionized first - OLTP or OLAP? Integration vs specialization, how to choose? Where will DBAs go in the AI era? Feng’s views from the HOW 2025 conference roundtable, organized and published.

The database for the AI era has been settled. Capital markets are making intensive moves on PostgreSQL targets, with PG having become the default database for the AI era.

Future software = Agent + Database. No middle tiers, just agents issuing CRUD. Database skills age well, and PostgreSQL is poised to be the agent-era default.