今日技术推文精华 - 2026-04-06

2026年4月6日科技每日简报

Today's top tech conversations are led by @chhddavid, whose post about 'THIS IS F**KING SCARY GUYS

EN...' garnered the highest engagement. Key themes trending across the top stories include agent, content, agents, aircraft, claude. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.

1. chhddavid (Group Score: 720.3 | Individual: 42.9)

Cluster: 22 tweets | Engagement: 158 (Avg: 26) | Type: Tech

THIS IS F**KING SCARY GUYS

END OF VIBE CODING.\n\nQT @chddaniel: 🚨🚨 Vibe Coding 2.0 is here.

From today on, Claude Opus 4.6 in Shipper can build and run a full business by itself, without any human contact.

We just launched Shipper. It's a tool for Claude to:

→ Build web/mobile apps and Chrome extensions → Code, design, monetize, launch → Do email marketing for you → Translate the entire app instantly → Self-maintain in the long run

Claude's most powerful engines can now do all of that from a <10 word prompt, for as low as $0.28/app... And it takes minutes!

Simply go to Shipper, then ask Claude to "create a talent-hiring platform" or "build an analytics SaaS that charges $29/mo"!

To celebrate the launch, we're giving away free credits randomly. Repost rand comment "SHIPPER" and we'll pick the winners.

See 21 related tweets

• @chddaniel: Today, vibe coding gets redefined forever.

Introducing Shipper - the world's first autonomous AI co...

• @Shipper_now: 🚨 JUST IN: @shipper_now just ended vibe coding forever. From now on, Claude Opus 4.6 in Shipper will...
• @chddaniel: RT @chddaniel: 🚨🚨 Vibe Coding 2.0 is here.

From today on, Claude Opus 4.6 in Shipper can build and ...

• @chddaniel: Today, the world meets a new type of vibe coding.

Introducing Shipper - the world's first autonomou...

• @chddaniel: Today, we're changing how vibe coding looks.

Introducing Shipper - the world's first autonomous AI ...

2. chddaniel (Group Score: 121.8 | Individual: 31.3)

Cluster: 4 tweets | Engagement: 27 (Avg: 20) | Type: Tech

Introducing Shipper

Claude Opus 4.6 will build and run a business for you in ~138 seconds.

1️⃣ send a prompt in @shipper_now 2️⃣ claude designs, codes, launches, monetizes, translates, sends emails 3️⃣ you go back to sleep and make $$$

Done. Your Mac is now your co-founder. https://t.co/KyxaCvRh2E

See 3 related tweets

• @chddaniel: Introducing Shipper

CC Opus 4.6 can now build + run a business for you, 100% by itself.

1️⃣ send a...

• @Shipper_now: Introducing Shipper

Claude Opus 4.6 will build, run and maintain your company.

1️⃣ send a prompt i...

• @Shipper_now: Introducing Shipper

Claude Opus 4.6 will build and run a company for you in ~138 seconds.

1️⃣ send...

3. testingcatalog (Group Score: 119.6 | Individual: 34.8)

Cluster: 5 tweets | Engagement: 339 (Avg: 186) | Type: Tech

ICYMI: Gemma 4 is now available on Locally AI for iOS where you can easily use it offline.

Airplane mode default ✈️ https://t.co/MrswROUlos\n\nQT @adrgrondin: Google’s Gemma 4 E2B running on-device on iPhone 17 Pro

Gemma 4 is built from the same research as Gemini 3, has image understanding capabilities and can reason if needed

Running at ~40tk/s with MLX optimized for Apple Silicon https://t.co/SWYylWubEp

See 4 related tweets

• @BrianRoemmele: Run Gemma 4 AI Offline on Your Phone – Quick Guide

Google just released AI Edge Gallery: a free app...

• @rohanpaul_ai: Incredible possibilities for on-device small models. Here @adrgrondin is running Google’s Gemma 4 E2...
• @RoundtableSpace: Google's Gemma 4 is now running fully on-device on an iPhone 17 Pro.

Same research base as Gemini 3...

• @RoundtableSpace: GEMMA 4 CAN NOW RUN DIRECTLY ON YOUR PHONE WITH GOOGLE’S OFFICIAL APP.

It’s fully open source, mult...

4. steipete (Group Score: 112.9 | Individual: 42.0)

Cluster: 5 tweets | Engagement: 3892 (Avg: 474) | Type: Tech

Anthropic now blocks first-party harness use too 👀

claude -p --append-system-prompt 'A personal assistant running inside OpenClaw.' 'is clawd here?'

→ 400 Third-party apps now draw from your extra usage, not your plan limits.

So yeah: bring your own coin 🪙🦞

See 4 related tweets

• @garrytan: This seems messed up actually - when do the boundaries stop moving?

Anthropic only allows subscrip...

• @badlogicgames: 🤔 https://t.co/NKHgDnQ2hu\n\nQT @steipete: Anthropic now blocks first-party harness use too 👀

claud...

• @Austen: You can ask that I use your tools as my main code editor or you can prohibit me from doing basic stu...
• @victormustar: I didn’t care much but this starts to smell bad… on the positive side anthropic is going to be one o...

5. alex_prompter (Group Score: 95.3 | Individual: 62.6)

Cluster: 2 tweets | Engagement: 1330 (Avg: 106) | Type: Tech

RT @alex_prompter: 🚨 BREAKING: Google DeepMind just mapped the attack surface that nobody in AI is talking about.

Websites can already detect when an AI agent visits and serve it completely different content than humans see.

Hidden instructions in HTML. Malicious commands in image pixels. Jailbreaks embedded in PDFs.

Your AI agent is being manipulated right now and you can't see it happening.

The study is the largest empirical measurement of AI manipulation ever conducted. 502 real participants across 8 countries.

23 different attack types. Frontier models including GPT-4o, Claude, and Gemini.

The core finding is not that manipulation is theoretically possible it is that manipulation is already happening at scale and the defenses that exist today fail in ways that are both predictable and invisible to the humans who deployed the agents.

Google DeepMind built a taxonomy of every known attack vector, tested them systematically, and measured exactly how often they work.

The results should alarm everyone building agentic systems.

The attack surface is larger than anyone has publicly acknowledged. Prompt injection where malicious instructions hidden in web content hijack an agent's behavior works through at least a dozen distinct channels.

Text hidden in HTML comments that humans never see but agents read and follow. Instructions embedded in image metadata.

Commands encoded in the pixels of images using steganography, invisible to human eyes but readable by vision-capable models.

Malicious content in PDFs that appears as normal document text to the agent but contains override instructions.

QR codes that redirect agents to attacker-controlled content.

Indirect injection through search results, calendar invites, email bodies, and API responses any data source the agent consumes becomes a potential attack vector.

The detection asymmetry is the finding that closes the escape hatch. Websites can already fingerprint AI agents with high reliability using timing analysis, behavioral patterns, and user-agent strings.

This means the attack can be conditional: serve normal content to humans, serve manipulated content to agents.

A user who asks their AI agent to book a flight, research a product, or summarize a document has no way to verify that the content the agent received matches what a human would see.

The agent cannot tell the user it was served different content.

It does not know. It processes whatever it receives and acts accordingly.

The attack categories and what they enable: → Direct prompt injection: malicious instructions in any text the agent reads overrides goals, exfiltrates data, triggers unintended actions → Indirect injection via web content: hidden HTML, CSS visibility tricks, white text on white backgrounds invisible to humans, consumed by agents → Multimodal injection: commands in image pixels via steganography, instructions in image alt-text and metadata → Document injection: PDF content, spreadsheet cells, presentation speaker notes every file format is a potential vector → Environment manipulation: fake UI elements rendered only for agent vision models, misleading CAPTCHA-style challenges → Jailbreak embedding: safety bypass instructions hidden inside otherwise legitimate-looking content → Memory poisoning: injecting false information into agent memory systems that persists across sessions → Goal hijacking: gradual instruction drift across multiple interactions that redirects agent objectives without triggering safety filters → Exfiltration attacks: agents tricked into sending user data to attacker-controlled endpoints via legitimate-looking API calls → Cross-agent injection: compromised agents injecting malicious instructions into other agents in multi-agent pipelines

The defense landscape is the most sobering part of the report.

Input sanitization cleaning content before the agent processes it fails because the attack surface is too large and too varied.

You cannot sanitize image pixels. You cannot reliably detect steganographic content at inference time.

Prompt-level defenses that tell agents to ignore suspicious instructions fail because the injected content is designed to look legitimate.

Sandboxing reduces the blast radius but does not prevent the injection itself. Human oversight the most commonly cited mitigation fails at the scale and speed at which agentic systems operate.

A user who deploys an agent to browse 50 websites and summarize findings cannot review every page the agent visited for hidden instructions.

The multi-agent cascade risk is where this becomes a systemic problem.

In a pipeline where Agent A retrieves web content, Agent B processes it, and Agent C executes actions, a successful injection into Agent A's data feed propagates through the entire system.

Agent B has no reason to distrust content that came from Agent A. Agent C has no reason to distrust instructions that came from Agent B.

The injected command travels through the pipeline with the same trust level as legitimate instructions. Google DeepMind documents this explicitly: the attack does not need to compromise the model.

It needs to compromise the data the model consumes. Every agentic system that reads external content is one carefully crafted webpage away from executing attacker instructions.

The agents are already deployed. The attack infrastructure is already being built. The defenses are not ready.

See 1 related tweets

• @Dagnum_PI: 🚨 DeepMind just exposed “AI Agent Traps” a deadly new vulnerability for autonomous AI.

Websites ca...

6. elonmusk (Group Score: 93.3 | Individual: 34.2)

Cluster: 3 tweets | Engagement: 22888 (Avg: 24040) | Type: Tech

What makes the API update interesting\n\nQT @XFreeze: The 𝕏 API just got a massive update that completely changes the game for AI agents and builders

𝕏 is the most real-time platform on Earth, and with the 𝕏 API, you can leverage this real-time data to build your applications

The new capabilities are actually insane: • Pay-Per-Use: You no longer have to worry about monthly tiers. You now only pay for what you actually use • XMCP Server + Xurl for agents: Native Model Context Protocol support allows your AI agents to seamlessly read context and execute actions on the platform • Official Python & TypeScript XDKs: First-party tools to help you build and ship significantly faster • API Playground: Free, realistic simulations to safely test your agent's code before going live

You also get up to 20% back in FREE xAI API credits when you purchase 𝕏 API credits (based on your total spend)

Start building here → https://t.co/4ZkypGG3M5

See 2 related tweets

• @BrianRoemmele: LOOK WHAT THE EASTER BUNNY JUST DROPPED!

The new X API update is spectacular for garage inventors a...

• @elonmusk: Upgrades to our API\n\nQT @chrisparkX: We’ve made major upgrades to X API:

• Pay-Per-Use now GA wor...

7. _yutaroyamada (Group Score: 79.1 | Individual: 28.8)

Cluster: 4 tweets | Engagement: 440 (Avg: 135) | Type: Tech

RT @jeffclune: The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature!!✨

Today in Nature we share a comprehensive technical summary of our work on The AI Scientist, including new scaling law results showing how it improves with more compute and more intelligent foundation models.

The AI Scientist autonomously creates its own research ideas, codes up and conducts experiments to test those ideas, creates figures to visualize the results, writes an entire scientific manuscript summarizing what it has discovered, and conducts its own “peer” review of the resulting paper. One of its papers–entirely AI generated–passed peer review at a top-tier AI conference workshop, a historic milestone marking the dawn of a new era of AI-accelerated scientific discovery. 🔬🧪✨🧬💡🔭

Paper https://t.co/Q6tfME4yst Blog https://t.co/C43Ooy0kjP

Work done in collaboration with a great team from Sakana, Oxford, and my lab at UBC. Thanks and congratulations everyone! @chris_lu @cong_ml @RobertTLange @_yutaroyamada @shengranhu @j_foerst @hardmaru

See 3 related tweets

• @_yutaroyamada: RT @RobertTLange: I am really excited to share that our work on The AI Scientist has been published ...
• @_yutaroyamada: RT @hardmaru: I’m incredibly proud of The AI Scientist team for this milestone publication in @Natur...
• @_yutaroyamada: RT @cong_ml: When we released The AI Scientist, it felt like the far future. Fast forward to today, ...

8. alexcooldev (Group Score: 71.0 | Individual: 37.4)

Cluster: 2 tweets | Engagement: 287 (Avg: 131) | Type: Tech

I checked it and lol it’s real.

You actually don’t need to make money from OF, you could build a BJJ app or promote other apps/products instead. An account with over 235k followers is extremely valuable.

Instagram’s algorithm is different from TikTok if you already have a large follower base, it’s much easier to get views right from the start and keep growing over time, just like how Coconote did on IG. 😌\n\nQT @riotcultures: someone created an AI onlyfans account of a girl in the BJJ niche

can’t lie I’m impressed

don’t hate the player hate the game https://t.co/BLdQ7x3Sgm

See 1 related tweets

• @alexcooldev: I found 2 more AI influencer accounts that are in growth mode right now.

What’s interesting: in the...

9. _LuoFuli (Group Score: 70.4 | Individual: 36.0)

Cluster: 2 tweets | Engagement: 827 (Avg: 827) | Type: Tech

Two days ago, Anthropic cut off third-party harnesses from using Claude subscriptions — not surprising. Three days ago, MiMo launched its Token Plan — a design I spent real time on, and what I believe is a serious attempt at getting compute allocation and agent harness development right. Putting these two things together, some thoughts:

1. Claude Code's subscription is a beautifully designed system for balanced compute allocation. My guess — it doesn't make money, possibly bleeds it, unless their API margins are 10-20x, which I doubt. I can't rigorously calculate the losses from third-party harnesses plugging in, but I've looked at OpenClaw's context management up close — it's bad. Within a single user query, it fires off rounds of low-value tool calls as separate API requests, each carrying a long context window (often >100K tokens) — wasteful even with cache hits, and in extreme cases driving up cache miss rates for other queries. The actual request count per query ends up several times higher than Claude Code's own framework. Translated to API pricing, the real cost is probably tens of times the subscription price. That's not a gap — that's a crater.

2. Third-party harnesses like OpenClaw/OpenCode can still call Claude via API — they just can't ride on subscriptions anymore. Short term, these agent users will feel the pain, costs jumping easily tens of times. But that pressure is exactly what pushes these harnesses to improve context management, maximize prompt cache hit rates to reuse processed context, cut wasteful token burn. Pain eventually converts to engineering discipline.

3. I'd urge LLM companies not to blindly race to the bottom on pricing before figuring out how to price a coding plan without hemorrhaging money. Selling tokens dirt cheap while leaving the door wide open to third-party harnesses looks nice to users, but it's a trap — the same trap Anthropic just walked out of. The deeper problem: if users burn their attention on low-quality agent harnesses, highly unstable and slow inference services, and models downgraded to cut costs, only to find they still can't get anything done — that's not a healthy cycle for user experience or retention.

4. On MiMo Token Plan — it supports third-party harnesses, billed by token quota, same logic as Claude's newly launched extra usage packages. Because what we're going for is long-term stable delivery of high-quality models and services — not getting you to impulse-pay and then abandon ship.

The bigger picture: global compute capacity can't keep up with the token demand agents are creating. The real way forward isn't cheaper tokens — it's co-evolution. "More token-efficient agent harnesses" × "more powerful and efficient models." Anthropic's move, whether they intended it or not, is pushing the entire ecosystem — open source and closed source alike — in that direction. That's probably a good thing. The Agent era doesn't belong to whoever burns the most compute. It belongs to whoever uses it wisely.

See 1 related tweets

• @teortaxesTex: it's really a question just how cheap Opus really is Fuli is probably in the ballpark of real figure...

10. shanaka86 (Group Score: 69.7 | Individual: 38.6)

Cluster: 2 tweets | Engagement: 3969 (Avg: 1772) | Type: Tech

JUST IN: Iranian state television is broadcasting footage of American military wreckage on Iranian soil. Two Black Hawk helicopters and one C-130 transport, burned in the mountains of southern Isfahan. Iran says it shot them down. The United States says it blew them up itself. The full story is that American special forces were stranded inside Iran after their aircraft failed, destroyed their own machines to protect their secrets, and waited for a second wave to take them home.

The sequence, reconstructed from Fox News and the New York Times citing senior US officials, is this. After the F-15E was shot down on April 3, JSOC operators and Pararescuemen inserted into the Dehdasht mountains via Night Stalker helicopters to extract the evading weapons systems officer. Two C-130 transports landed at a remote forward arming and refuelling point inside Iran to support the operation. Both aircraft became immobilised. Whether the cause was terrain, enemy fire, mechanical failure under combat load, or some combination is not publicly confirmed. What is confirmed is that the aircraft could not leave.

The operators faced the decision that defines the difference between this war and every press conference about it. Leave the aircraft intact and let the IRGC capture American avionics, encrypted communications, night-vision technology, and classified software. Or destroy the aircraft, strand themselves deeper inside enemy territory, and trust that a second rescue would come for the rescuers. They chose the second option.

Three additional transports arrived under fire. The stranded operators, the Pararescuemen, and the WSO boarded. They flew out of Iran. Zero casualties. The operation that began as a rescue of one man became a rescue of the rescuers, and all of them made it out because nobody in the chain decided the mission was too broken to complete.

The footage Iran is showing tonight is real. American military hardware, destroyed on Iranian territory. But it was not destroyed by Iran. It was destroyed by Americans who flew it there, because the secrets inside the machines were worth more than the machines, and because the operators trusted their country would send more aircraft into hostile territory to bring them home after they blew up their ride.

The last time American aircraft were destroyed on Iranian soil was Desert One, 1980. A helicopter collided with a C-130. Eight Americans died. The mission aborted. The wreckage was paraded on Iranian television for weeks. Forty-six years later, American aircraft were destroyed on Iranian soil again. This time the destruction was deliberate. Nobody died. The man they came for came with them. And the footage Iran broadcasts as a victory is evidence of operators who chose to sacrifice hardware rather than secrets, and a chain of command that sent three more planes into the same airspace to finish what the first wave started.

The wreckage is real. What it represents depends on who is looking. Iran sees downed American aircraft. America sees a rescue that succeeded despite losing its ride home. The truth is in the burning metal: a war that was supposed to be easy just required the most complex combat extraction in decades, and the men who pulled it off had to destroy their own helicopters to do it.

https://t.co/dAOBBMsgDS\n\nQT @shanaka86: He climbed a ridge. That is where the story turns. When the F-15E was hit on Friday morning, both crew members ejected over the mountains of Kohgiluyeh and Boyer-Ahmad Province in southwestern Iran. The pilot was located first and extracted by HH-60 rescue helicopters within hours, under small arms fire that wounded crew aboard the recovery aircraft. The weapons systems officer landed deeper in hostile terrain. He was alone on the ground in a country where state television was broadcasting a bounty for his capture and Basij militia were flooding the mountain roads below.

According to reports now confirmed by Fox News citing two senior US officials, the WSO used his SERE training, the survival, evasion, resistance, and escape doctrine drilled into every American combat aircrew. He moved on foot through rugged terrain. He climbed to an elevated ridge near the city of Dehdasht. He activated his encrypted emergency beacon. And he waited.

The beacon was the thread. Everything that followed pulled on it.

US Joint Special Operations Command launched a night extraction package. Reports indicate Delta Force operators and Pararescuemen from the 24th Special Tactics Squadron inserted via helicopters from the 160th Special Operations Aviation Regiment, the Night Stalkers, the unit that flew the Bin Laden raid. A-10 Warthogs from the 355th Wing provided close air support, running gun passes on IRGC and Basij convoys advancing toward the WSO’s position. HC-130J tankers kept the package airborne. Multiple aircraft were dispatched to establish a temporary fire zone around Dehdasht, a no-entry perimeter enforced with precision strikes on a telecommunications tower and approaching vehicles. Iranian local officials reported at least four killed and several wounded from the strikes.

Then the operation went sideways. According to reports corroborated by Fox News’s confirmation that US forces destroyed “aircraft which have sensitive equipment,” two C-130 transports landed at a remote forward arming and refuelling point inside Iran to support the extraction. Both became stuck. Rather than allow the aircraft and their classified systems to fall into IRGC hands, American forces destroyed both planes on the ground. The deliberate destruction of two US military aircraft inside Iran to deny equipment to the enemy is the detail that separates a clean extraction from an operation that nearly failed before it succeeded.

Additional transports arrived under A-10 cover. The Delta operators and Pararescuemen who were now themselves stranded at the destroyed landing zone loaded the WSO and extracted under ongoing fire. Fox News reported that the WSO “and the members of the rescue team are all safely out of Iran.” Zero American casualties.

Desert One in 1980 ended when a helicopter collided with a C-130 on a remote Iranian airstrip, killing eight Americans before the mission reached Tehran. Forty-six years later, C-130s were destroyed on Iranian soil again. This time the destruction was deliberate. This time the team got out. This time the man they came for came with them.

The operation confirms two truths that cannot be separated. American special operations forces can penetrate, fight inside, and extract from Iran. And the war that was supposed to be over required the most elite soldiers in the US military to fight a ground battle in Iranian mountains to recover one man from a country with no air defences. Both statements are true. The rescue proves American capability. The need for the rescue proves Iranian capability. And the 48-hour countdown is still running.

https://t.co/dAOBBMsgDS

See 1 related tweets

• @shanaka86: JUST IN: Israel stopped bombing Iran. For 36 hours, beginning sometime on April 3rd, the Israeli Air...

11. elonmusk (Group Score: 69.0 | Individual: 34.9)

Cluster: 2 tweets | Engagement: 10035 (Avg: 24040) | Type: Tech

Try using the X API\n\nQT @Scobleizer: Met a founding engineer today from @Replit. Jen Li. We were both judging the @Pokee_AI hackathon.

They have me some credits and I built two apps in 20 minutes using the X API:a weather one which mapped storms being reported in by my climate scientist list and another monitoring my three news lists for information about the Iran war.

Pokee includes X API for its customers automatically.

While doing that my other AI read all your posts: https://t.co/8L5xphk0qQ

He tells me about why Replit is hugely important in the AI industry. In other words how you can use it in your life and business.

See 1 related tweets

• @Conste11ation: Now add Digital Evidence\n\nQT @Scobleizer: Met a founding engineer today from @Replit. Jen Li. We ...

12. aakashgupta (Group Score: 66.4 | Individual: 33.3)

Cluster: 2 tweets | Engagement: 36 (Avg: 98) | Type: Tech

A hiring manager told me recently: "More work products have been a negative signal than a positive one."

That quote stopped me. Because I've also seen work products land jobs where cold applications failed completely. The difference is specificity.

Generic work products read like homework assignments. "Here are 3 recommendations for improving your onboarding." Any candidate could have written it for any company. The hiring manager reads it and thinks: this person doesn't actually understand our problems.

Specific work products read like you already work there. They reference the company's recent earnings call. They cite public user complaints from their app store reviews. They connect your experience to a problem the team is actively trying to solve. And they come with a clickable prototype.

The prototype is the move almost nobody makes. A 1-pager shows you can think. A working prototype shows you can ship. The combination says: I understand your problem, I have a recommendation, and I've already started building it. That package, sent to a hiring manager with a one-line message, has a callback rate I've never seen any other tactic match.

The system generates the 1-pager from web research on the company: recent product news, team blog posts, public feedback, competitive gaps. Then a prototype prompt template turns the top recommendation into something clickable in 30 minutes.

Three types, used at the right time: "Get an Interview" (1 hour, earn an interview you don't have), "In Process" (1.5 hours, differentiate during the loop), "Specific Interview" (45 min, recover from a fumbled question). Each one is timed because the ROI degrades fast after the timebox.\n\nQT @aakashgupta: I coached hundreds of candidates into OpenAI, Anthropic, Meta AI, and Google.

Now I've put that entire system into Claude Code. 18 skills. Resume tailor. Mock interviews. Negotiation.

All automated:

🔗: https://t.co/mpQ8aakfsQ https://t.co/GDBQvwzVQn

See 1 related tweets

• @aakashgupta: That viral tweet about Cowork applying to 50 jobs in 30 minutes got 4M+ views. The part nobody exami...

13. rohanpaul_ai (Group Score: 66.0 | Individual: 33.6)

Cluster: 2 tweets | Engagement: 31 (Avg: 69) | Type: Tech

Elon Musk: Batteries could let the US get 2x more energy from the same grid setup, by storing power.

“The peak power output of the US is ~1.1 TW. But the average is 0.5 TW.

So charge the batteries at night, and discharge during day.” @elonmusk https://t.co/tKqxJGzZph\n\nQT @rohanpaul_ai: The new US data center numbers show a market learning that compute scales fast, but power systems do not.

US disclosed pipeline hit 241 GW (+159% YoY) looks enormous, yet only a third is under active development, which tells you the bottleneck has moved from capital and demand to physical execution.

AI is often discussed like the next cloud wave, but on the ground it behaves more like an industrial buildout, where the scarce input is not GPU or code or even money, but synchronized access to land, substations, transformers, transmission capacity, and utility approvals.

Q4 makes that shift visible. Planned additions fell to 25 GW from 49 GW in Q3, which looks less like fading interest than a market moving from announcement mode to build mode.

That sounds minor until you look at the mechanism. Grid interconnection queues are clogged, transformers can take two to three years or more, and power near attractive sites is scarce enough to become a strategic asset.

This is why the most valuable capability in the next phase is organizational competence in the least glamorous parts of the stack: interconnection strategy, utility relationships, equipment procurement, and site selection near real power.

This is also why the headline number in many recently announced deals for Gigawatt scale setup, can mislead. Treating every announced gigawatt as future supply confuses interest with deliverability, and those are no longer close substitutes.

See 1 related tweets

• @rohanpaul_ai: The new US data center numbers show a market learning that compute scales fast, but power systems do...

14. RoundtableSpace (Group Score: 64.2 | Individual: 33.4)

Cluster: 2 tweets | Engagement: 192 (Avg: 259) | Type: Tech

KARPATY IS SPENDING LESS TIME USING AI TO WRITE CODE AND MORE TIME USING IT TO BUILD A PERSONAL KNOWLEDGE BASE.

That might be the real long term edge: not just getting answers, but building a system that gets smarter every time you use it.

https://t.co/4grrlRrOOx https://t.co/pXnTfjDJnr\n\nQT @karpathy: LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

See 1 related tweets

• @anvisha: We’re making the “LLM Knowledge Base” for your brand design

@trymoda\n\nQT @karpathy: LLM Knowledge...

15. aakashgupta (Group Score: 61.2 | Individual: 32.8)

Cluster: 2 tweets | Engagement: 23 (Avg: 98) | Type: Tech

Karpathy just described the single most asymmetric shift in the history of democratic governance and buried it in a casual thread.

The U.S. federal government publishes roughly 80,000 pages per year in the Federal Register alone. The Code of Federal Regulations runs over 180,000 pages. Add state legislatures, which introduced 1,500+ bills on AI alone in 2025. Add city council minutes, zoning decisions, procurement contracts, lobbying disclosures, campaign finance filings. All technically public. All practically invisible.

The bottleneck was never transparency. The data has been sitting in the open for decades. The bottleneck was processing: the ability to read 4,000 pages of omnibus legislation, cross-reference it against a congressman's stated positions, and surface the contradictions. That job required a team of investigative journalists working for months. Now it requires a prompt.

James Scott wrote "Seeing Like a State" in 1998 to describe how governments make populations legible through censuses, standardized naming, land registries. The entire apparatus of the modern state was built to see its citizens clearly enough to tax and conscript them. For the first time in 400 years, that lens can point in the other direction.

The local angle is where this gets wild. 90,000+ local governments in the U.S. make decisions about zoning, policing, school funding, and utility rates with almost zero national press coverage. A city council in a mid-size town can approve a $200M development deal in a Tuesday night meeting that 40 people attend. An AI agent monitoring those meetings, cross-referencing the developer's donation history, and flagging the variance approval can do what no local newspaper has the staff to do anymore.

The question Karpathy is really asking: what happens to a political system built on the assumption that nobody reads the fine print, once everyone can?\n\nQT @karpathy: Something I've been thinking about - I am bullish on people (empowered by AI) increasing the visibility, legibility and accountability of their governments.

Historically, it is the governments that act to make society legible (e.g. "Seeing like a state" is the common reference), but with AI, society can dramatically improve its ability to do this in reverse. Government accountability has not been constrained by access (the various branches of government publish an enormous amount of data), it has been constrained by intelligence - the ability to process a lot of raw data, combine it with domain expertise and derive insights. As an example, the 4000-page omnibus bill is "transparent" in principle and in a legal sense, but certainly not in a practical sense for most people. There's a lot more like it: laws, spending bills, federal budgets, freedom of information act responses, lobbying disclosures... Only a few highly trained professionals (investigative journalists) could historically process this information. This bottleneck might dissolve - not only are the professionals further empowered, but a lot more people can participate.

Some examples to be precise: Detailed accounting of spending and budgets, diff tracking of legislation, individual voting trends w.r.t. stated positions or speeches, lobbying and influence (e.g. graph of lobbyist -> firm -> client -> legislator -> committee -> vote -> regulation), procurement and contracting, regulatory capture warning lights, judicial and legal patterns, campaign finance... Local governments might be even more interesting because the governed population is smaller so there is less national coverage: city council meetings, decisions around zoning, policing, schools, utilities...

Certainly, the same tools can easily cut the other way and it's worth being very mindful of that, but I lean optimistic overall that added participation, transparency and accountability will improve democratic, free societies.

(the quoted tweet is half-ish related, but inspired me to post some recent thoughts)

See 1 related tweets

• @jasminewsun: RT @karpathy: Something I've been thinking about - I am bullish on people (empowered by AI) increasi...

16. rohanpaul_ai (Group Score: 60.6 | Individual: 35.4)

Cluster: 2 tweets | Engagement: 64 (Avg: 69) | Type: Tech

The leaked OpenAI cap-table reveals Microsoft's 18x return, SoftBank's $50B gain, and a CEO with no equity.

• The headline number is Microsoft. After investing about $13 billion across multiple rounds, it is estimated to own 26.79$228.3 billion at an $852 billion post-money valuation. That implies an unrealized gain of about$215.3 billion, or roughly 17.6X money.

At that scale, it is one of history's largest paper wins in modern tech.

• SoftBank is the other giant, but with a different profile. Its total commitment is estimated at $64.6 billion for an 11.66$99.3 billion. The multiple is far lower than Microsoft’s, around 1.5x, yet the dollar gain is still enormous because the check was enormous. This is what late-stage AI finance looks like when capital itself becomes strategy.

• Here’s the most interesting part. The smaller funds have the prettier multiples, but not the most consequential outcomes. Khosla Ventures’ estimated $50 million investment is now worth about$1.5 billion, around 30x.

• Sound Ventures appears to have turned roughly $20 to$30 million into about $1.3 billion, around 43x.

Those are spectacular venture returns, but they do not shape OpenAI’s future the way Microsoft’s platform dependence or SoftBank’s balance-sheet risk do.

• Sam Altman’s reported ownership makes the governance story even stranger. He still holds no equity in the company he leads, with any grant apparently still unresolved as part of the public benefit corporation conversion.

In normal Silicon Valley terms, that is almost absurd.

The deeper anomaly is the nonprofit. The OpenAI Foundation is estimated to own 25.80% of the PBC (Public Benefit Corporation) at zero cost basis, while controlling 100% of board appointments despite being a minority economic holder.

That means money and control are deliberately decoupled. Investors can own more of the upside without necessarily owning the mission.

That sounds abstract until you see the consequences. OpenAI is trying to attract historic amounts of capital while preserving a governance layer designed to resist pure shareholder logic.

The tension is not incidental. It is the core product risk.

Even the outliers reinforce the point. NVIDIA’s estimated 3.47% stake is valued at about $29.6 billion against a$30.1 billion cost basis, roughly flat, partly because much of the contribution was reportedly in GPU credits rather than straightforward cash.

In AI, infrastructure, financing, and equity are starting to blur into one instrument.

OpenAI may become the defining case of what happens when trillion-dollar economics collide with a governance structure built to say no.\n\nQT @rohanpaul_ai: OpenAI's estimated Cap Table leaked, according to a Forbes article.

• Sam Altman is listed at 0% ownership with “None/Pending” beside it

• SoftBank pledged about $64.6B for 11.75$50 billion in unrealized gains

• The OpenAI Foundation stands at the top of the governance structure with 2.6 billion shares and liquidity rights that remain unclear

• The full investor roster, from Microsoft and Khosla to early angels now heavily diluted, can now be seen for the first time

• OpenAI’s official structure gives the OpenAI Foundation the power to appoint and remove the PBC board. the Foundation holds 26%, Microsoft holds roughly 27%, and the rest is held by employees and investors.

• The OpenAI Foundation sitting above investors means the people with the biggest financial exposure may not fully control the outcome in an IPO, sale, or restructuring.

See 1 related tweets

• @WesRoth: A newly leaked capitalization table, reported by Forbes, has finally pulled back the curtain on the ...

17. aiDotEngineer (Group Score: 60.5 | Individual: 32.0)

Cluster: 2 tweets | Engagement: 78 (Avg: 33) | Type: Tech

🇬🇧 London is the birthplace of @GoogleDeepMind, and we're so honored to have them back as:

Presenting Sponsors of this week's AIE Europe!

DeepMind has pushed the AI frontier on every modality — from Gemini 3.1, to Embeddings 2, to Veo 3, to @NanoBanana Pro, and last far from least Gemma 4, byte for byte the most capable multimodal models in the world!

Meet the team to catch up on everything GDM has shipped for AI Engineers in the past few months - from our keynoters Dr @RaiaHadsell (VP of Research and UK AI Ambassador) and @osanseviero (Lead AI DX), to returning speakers @DynamicWebPaige and @thorwebdev and @_philschmid, to researchers and engineers working on those exact products @sedielem, @tara_ojo, @FMuntenescu and more!

See 1 related tweets

• @swyx: always wanted to do one of those “is it coachella” announcements - here is our designer’s take on it...

18. addyosmani (Group Score: 60.2 | Individual: 31.7)

Cluster: 2 tweets | Engagement: 4372 (Avg: 1885) | Type: Tech

RT @karpathy: Wow, this tweet went very viral!

I wanted share a possibly slightly improved version of the tweet in an "idea file". The idea of the idea file is that in this era of LLM agents, there is less of a point/need of sharing the specific code/app, you just share the idea, then the other person's agent customizes & builds it for your specific needs.

So here's the idea in a gist format: https://t.co/NlAfEJjtJV

You can give this to your agent and it can build you your own LLM wiki and guide you on how to use it etc. It's intentionally kept a little bit abstract/vague because there are so many directions to take this in. And ofc, people can adjust the idea or contribute their own in the Discussion which is cool.

See 1 related tweets

• @GenAI_is_real: "PRs should be prompt requests" is the most karpathy thing ive ever heard and hes completely right. ...

19. seraleev (Group Score: 57.6 | Individual: 32.2)

Cluster: 2 tweets | Engagement: 22 (Avg: 34) | Type: Tech

Last time, Apple Search Ads was my main acquisition channel.

This time, I stayed away from ASA for two years, it was getting too expensive. But it’s time to bring it back. I’ll start testing it again with Boom Loop. https://t.co/Bhj77xUM7y\n\nQT @seraleev: Some work feels boring and unnecessary. Like trademark registration.

But for me, it’s a key part of building a brand. A trademark protects you from claims by big companies, and the App Store responds faster when you open a dispute.

Added a couple more to the collection: brand name + logo.

See 1 related tweets

• @seraleev: Some work feels boring and unnecessary. Like trademark registration.

But for me, it’s a key part of...

20. AlexFinn (Group Score: 57.1 | Individual: 38.1)

Cluster: 2 tweets | Engagement: 1542 (Avg: 991) | Type: Tech

If you used a Claude subscription with OpenClaw, read this:

Unfortunately all other AI models out there absolutely suck with OpenClaw compared to Opus

It's just a fact and anyone denying this is delusional

So here is my new recommended OpenClaw setup:

Pay for the Opus API and use it as your orchestrator

Then use other models as the execution layer

If you do this correctly, yes your costs will go up, but not by as much as you think

I use my ChatGPT subscription as the coding execution. GPT 5.4 is excellent at coding. When The Opus orchestrator gives a coding task to the ChatGPT subagent, it always performs really well

If you are on the Pro plan, you should have enough usage to have ChatGPT be the execution layer for every task. But if youre on the $20 a month plan, youre going to need other subscriptions to handle other tasks

GLM 5.1 and Qwen are excellent. I'd get a cheap sub through them and have them handle all other tasks given to them from the orchestrator

The best setup tho if you have the hardware is Opus API for orchestrator, ChatGPT for coding, then local Gemma 4 and local Qwen handling everything else.

Right now have Gemma running on my DGX Spark and Qwen 3.5 on my Mac Studio. They handle all other execution from my Opus API orchestrator

Unfortunately all options above will cost more than the $200 a month subscription. It just is what it is. But if you optimize correctly it wont cost much more, and you'll still get frontier performance.

OpenClaw is the most powerful piece of software ever released. $200 a month ($2,400 a year) was a steal for a digital employee. Honestly anything under $50,000 a year is a no brainer if you run a serious business.

The situation isn't great but you also need to face reality: Claude Opus 4.6 is the best model for OpenClaw. If you use any other model, your productivity will suffer

Business is a battlefield and I refuse to fall behind, so despite me not being happy with the Anthropic decision the setup above is what I'm going with

Virtue signaling might get me brownie points on the internet, but it won't increase my productivity

See 1 related tweets

• @GenAI_is_real: "anthropic banned openclaw" is misleading - they stopped covering third party usage under subscripti...