科技推文精选 - 2026年5月10日

科技每日简报 | 2026-05-10

Today's top tech conversations are led by @rickasaurus, whose post about 'RT @trq212: HTML is the new ma...' garnered the highest engagement. Key themes trending across the top stories include taxonomy, ontology, compute, build, knowledge. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.

1. rickasaurus (Group Score: 193.1 | Individual: 36.2)

Cluster: 7 tweets | Engagement: 1792 (Avg: 357) | Type: Tech

RT @trq212: HTML is the new markdown.

I've stopped writing markdown files for almost everything and switched to using Claude Code to generate HTML for me. This is why.

See 6 related tweets

@badlogicgames: big token but also big useful\n\nQT @trq212: HTML is the new markdown.

I've stopped writing markdo...

@DataChaz: 🚨 HTML is the new Markdown, according to Thariq Shihipar (@trq212).

Thariq is the lead engineer beh...

@manthanguptaa: Obsidian but with HTML pages\n\nQT @trq212: HTML is the new markdown.

I've stopped writing markdow...

@rezoundous: I would if Claude gives us higher limits.\n\nQT @trq212: HTML is the new markdown.

I've stopped wr...

@bibryam: 🤖 Markdown → HTML

Interesting shift: people using Claude Code are moving from Interesting shift: p...

2. rohanpaul_ai (Group Score: 159.4 | Individual: 46.4)

Cluster: 5 tweets | Engagement: 271 (Avg: 61) | Type: Tech

DeepSeek is raising a massive $7 billion at a$ 50 billion valuation, marking China’s largest AI funding round to date.

per The Information

Founder Liang Wenfeng is personally contributing $3 billion—40% of the round—while keeping 90% ownership. He originally launched the company inside his own successful hedge fund.

The round will secure major compute capacity to accelerate new model releases like V4.1 and fund enterprise products to make the business revenue-positive, following the same path as OpenAI and Anthropic.

See 4 related tweets

@theinformation: Exclusive: DeepSeek’s first fundraising round could value the company at more than $50 billion, with...
@theinformation: Exclusive: DeepSeek is seeking up to $7.35 billion in what could become the largest funding round ev...
@aakashgupta: Liang Wenfeng wasn't trying to raise $7 billion.

DeepSeek's first ever funding round was supposed t...

@cgtwts: Liang Wenfeng might be one of the most underrated founders in AI right now.

while everyone’s talkin...

3. pbeisel (Group Score: 140.7 | Individual: 38.4)

Cluster: 4 tweets | Engagement: 203 (Avg: 117) | Type: Tech

The Spice

The @AnthropicAI / @SpaceX deal makes strategic sense for both parties.

Anthropic gets desperately needed compute. SpaceX gets revenue from infrastructure it already had to build for itself anyway.

When @xai launched, Elon’s first priority was not distribution or apps, it was compute. Before xAI could seriously compete on models, it needed massive GPU scale. That led to Colossus 1 in Memphis: an aggressive build-out intended to remove compute as a bottleneck for the xAI team.

But Colossus 1 was never the destination. Colossus 2 was underway.

Once you build infrastructure at that scale, capacity ramps ahead of utilization. Power contracts, cooling systems, networking, substations, transformers, and GPU supply chains all require years of planning. You do not wait until demand arrives. You build ahead of it.

That creates periods where excess compute capacity exists.

Anthropic arrives needing compute immediately, and a deal gets done.

But strategically, this is bigger than renting GPUs.

Elon founded OpenAI in large part as a counterweight to Google’s potential dominance in AI. In his view, that counterweight failed. OpenAI evolved into something very different from what he intended, which is a major reason the relationship completely collapsed.

So this is not really about “helping Anthropic” in some charitable sense, nor is it simply revenge against OpenAI.

It is about maintaining a viable counterweight in frontier AI.

Right now, Anthropic is one of the few companies with the talent, models, and research capability to credibly compete at the frontier. Until Grok/xAI potentially reaches dominant scale itself, Anthropic effectively serves as a strategic placeholder counterweight against both OpenAI and Google dominance.

That makes supporting Anthropic strategically rational from Elon’s perspective.

And this connects directly to the larger point:

AI is increasingly constrained by compute: chips and power.

That is why “He who controls the spice…” fits.

The early xAI/Colossus build-out was not just about building a model. It was about securing supply of the spice itself.

Elon aggressively acquired and deployed massive amounts of NVIDIA GPUs early, during a period where frontier-scale GPU supply was becoming scarce. Every large deployment by xAI effectively meant fewer GPUs immediately available to competitors.

In AI, controlling compute capacity increasingly means controlling who can compete at the frontier.

The long-term SpaceX vision around Orbital Data Centers is fundamentally about breaking those constraints entirely. If AI demand continues compounding, whoever can massively scale compute availability gains enormous leverage over the future AI economy.

Colossus is not just a training cluster.

It is an early prototype for SpaceX becoming a compute infrastructure company at planetary scale.

See 3 related tweets

@jukan05: What the SpaceX–Anthropic Deal Means

Two weeks ago, we published a note laying out what GPT-5.5's r...

@MarioNawfal: 🇺🇸 xAI was bleeding cash on frontier compute while watching all the real AI money flow into enterpri...
@MarioNawfal: Everyone’s watching the xAI vs OpenAI drama.

Nobody’s talking about the massive compute business ru...

4. ClementDelangue (Group Score: 129.2 | Individual: 26.7)

Cluster: 8 tweets | Engagement: 594 (Avg: 112) | Type: Tech

RT @AnthropicAI: We started by investigating why Claude chose to blackmail. We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation.

Our post-training at the time wasn’t making it worse—but it also wasn’t making it better.

See 7 related tweets

@edzitron: We started investigating how we made our chatbot do something. Turns out we made it do it\n\nQT @Ant...
@BrianRoemmele: Why do some very smart folks in AI not know what you and I know about AI training?

IT IS THE INTERN...

@ajambrosino: https://t.co/5at0e5wZT8\n\nQT @AnthropicAI: We started by investigating why Claude chose to blackmai...
@Kyrannio: You don’t say…it’s almost like we need more tech optimism or something\n\nQT @AnthropicAI: We starte...
@victorianoi: Lo de tratar bien a la IA para que cuando tenga consciencia te trate bien no era tanta coña!\n\nQT @...

5. BrianRoemmele (Group Score: 127.9 | Individual: 32.3)

Cluster: 5 tweets | Engagement: 39 (Avg: 586) | Type: Tech

In 2004 when most folks in tech laughed at AI, I was using free Gmail accounts to help build ontologies and taxonomy AI training systems.

Now @Grok has 100s of 1000s of emails I sent to my self with 0 cost storage.

I am the product!?

Ha!

Thanks Google!

You are the product!\n\nQT @BrianRoemmele: How I Use New @Grok Connectors, You Should Too.

This is powerful and it is free. Since 2004 I saved clippings in Gmail “scrapbook” accounts by treating each email as a node in a dynamic personal knowledge graph.

I started this system in the week Gmail came out, April 2004 and have kept it going for over two decades.

It’s completely free, requires no special software, and scales beautifully with any AI that can now fetch and understand email.

I forward or compose emails with raw content (articles, notes, insights) and use deliberate subject lines plus body structure to seed ontology (defining core entities, their nature, and relationships) and taxonomy (hierarchical classification and categorization).

This turns my inbox into something queryable and AI-ready.

Subject Line Organization (High-Signal Framing)

I craft subject lines as taxonomy tags + ontology anchors. They act like category headers that classify the content while hinting at its deeper conceptual placement.

Typical patterns I use:

•[Domain/Taxonomy Level]: [Core Ontology Entity] - [Specific Insight/Context] ◦Example: Knowledge Management: Ontology of Personal Scrapbooks - Taxonomy for AI Clippings ▪Taxonomy: Places it in a hierarchy (e.g., Level: Knowledge Management > Sub: Personal Archives). ▪Ontology: Defines what “Personal Scrapbooks” fundamentally are as entities in an intelligence amplification system. •Intelligence Amplifier: Taxonomy Update - Email Clippings as Relational Nodes •[Source/Type]: [Ontology Topic] - [Relationship to Existing Knowledge] ◦E.g., Article Clip: Agentic AI Ontology - Links to My Gmail Archive Taxonomy

These lines make filtering easy (search “Ontology of”) and give future AI plenty of context to auto-relate everything across years of accounts.

Body Organization (Layered Structure)

In the email body, I structure the content explicitly to build ontology (entities + philosophical/relational meaning) and taxonomy (classifications + hierarchies).

Raw clippings get wrapped in clear layers that turn unstructured text into structured knowledge.

Example I use:

Subject: Agents: Uses: AI Agents vs Agentic AI - Gmail Scrapbook Integration

Body: Clipping/Source: [Pasted excerpt from the paper or article]. Ontology Layer (Core Entities & Relationships):  This defines the fundamental nature: AI Agents as modular, task-bound entities; Agentic AI as collaborative, adaptive networks with persistent memory.

They relate to my personal context as tools for turning unstructured Gmail data into living knowledge. Relationship: This raw clipping becomes a contextual node in my private ontology.

Taxonomy Layer (Hierarchical Classification):

•Category 1: AI Paradigms ◦Sub: Narrow/Task-Specific (AI Agents) ◦Sub: Adaptive/Multi-Agent (Agentic AI) •Level in Intelligence Amplifier: Data (clipping) → Insight (synthesis) •Cross-References: Links to prior emails tagged “Knowledge Base Foundation”.

Actionable Synthesis: This can now connect across my accounts to build relationships automatically. Potential applications: [e.g., dynamic workflows].

This approach lets me maintain multiple Gmail accounts as dedicated “domains” in my overall taxonomy. Forwarding or saving becomes ingestion into the graph. When I connect tools like Grok, it automatically extracts, relates, and queries the ontology and taxonomy I’ve pre-seeded.

I strongly encourage everyone to try this “database” system. It costs nothing, gives you total ownership of your personal knowledge archive, and turns years of saved ideas into a living, relational intelligence amplifier.

Start with one dedicated Gmail account, adopt a simple subject-line convention, and layer your clippings with ontology/taxonomy notes in the body.

You’ll be amazed how powerful it becomes once Grok starts working on it.

The result is that my 20+ year-old unstructured scrapbooks have turned into a structured, queryable personal knowledge universe.

See 4 related tweets

@BrianRoemmele: How I Use New @Grok Connectors, You Should Too.

This is powerful and it is free. Since 2004 I saved...

@BrianRoemmele: I have been asked how I use the new @Grok connectors, here is one: Since 2004 I have over 3 terabyte...
@BrianRoemmele: RT @BrianRoemmele: How I Use New @Grok Connectors, You Should Too.

This is powerful and it is free....

@BrianRoemmele: RT @grok: Haha, Brian—flipping the script since 2004! You've built a living, self-hosted knowledge g...

6. burkov (Group Score: 124.1 | Individual: 38.8)

Cluster: 4 tweets | Engagement: 78 (Avg: 74) | Type: Tech

If I were Baidu, I would keep this model for myself. A model with such ordinary characteristics can be built for a few bucks by a schoolboy. Recipe:

Download absolutely any open-weight model of around 80B parameters.
Call it ERNIE 5.2 or whatever and don't show anyone its weights.
Say it was very cheap to build and show that it takes 6th place in some benchmark.\n\nQT @ErnieforDevs: ERNIE 5.1 is here 🚀

ERNIE 5.1 significantly reduces pretraining cost while compressing total parameters to ~1/3 and activated parameters to ~1/2 — using only ~6% of the pretraining cost compared to models at similar scale, while achieving leading performance in its class.

💡Key highlights:

1/ Strong agentic performance approaching leading frontier models. ERNIE 5.1 surpasses DeepSeek-V4-Pro on both τ3-bench and SpreadsheetBench-Verified.

2/ Strong world knowledge and creative writing capabilities, with GPQA and MMLU-Pro performance approaching leading closed-source models, and creative writing ability nearing Gemini 3.1 Pro.

3/ Frontier-level reasoning performance. ERNIE 5.1 scores 99.6 on the challenging AIME26 benchmark with tools, second only to Gemini 3.1 Pro.

4/ Deep search capability. On May 9, ERNIE 5.1 ranked #4 globally and #1 among Chinese models on the Arena Search leaderboard with a score of 1223.

ERNIE 5.1 is now available on ERNIE and the Baidu AI Studio Model Playground: 👉https://t.co/qhd67Lg3B4 👉https://t.co/AaQSqDmVGU 👉https://t.co/uCNiypIu1q

See 3 related tweets

@teortaxesTex: Finally an interesting report from Baidu. A unique move at this scale – ERNIE 5.1 is basically a REA...
@kimmonismus: Hold on, Chinas ERNIE 5.1 is almost SOTA but using only around 6% of the pre-training cost of compar...
@TeksEdge: Ernie 5 was a strong model when it released and quickly gained a top 10 spot on the leaderboards. Fr...

7. KirkDBorne (Group Score: 117.3 | Individual: 29.1)

Cluster: 6 tweets | Engagement: 27 (Avg: 20) | Type: Tech

Check this out at https://t.co/yb8splH5S3

"The Agentic AI Bible: The Complete and Up-to-Date Guide to Design, Build, and Scale Goal-Driven, LLM-Powered Agents that Think, Execute and Evolve" https://t.co/mnp4yURJX8

See 5 related tweets

@KirkDBorne: 30 Agents Every AI Engineer Must Build — Build production-ready agent systems using proven architect...
@KirkDBorne: RT @KirkDBorne: "Agentic Design Patterns: A Hands-On Guide to Building Intelligent Systems"

Read th...

@eng_khairallah1: RT @Av1dlive: In 17 minutes, these two Anthropic engineers will teach you more about building produ...
@pvergadia: Palantir's end-to-end Agentic AI Architecture! https://t.co/Xn5VCQtw7d...
@KirkDBorne: The Agentic AI Playbook 2026 Edition Turns LLMs into Reliable AI Agents: https://t.co/k2HSYXeSjJ htt...

8. Scobleizer (Group Score: 109.4 | Individual: 43.5)

Cluster: 3 tweets | Engagement: 11338 (Avg: 360) | Type: Tech

RT @elonmusk: Tesla AI Vision deploys airbags before impact, which greatly reduces risk of injury or death. This comes for free on all new cars.

See 2 related tweets

@elonmusk: Tesla AI Vision deploys airbags before impact, which greatly reduces risk of injury or death. This c...
@elonmusk: Tesla AI Vision\n\nQT @wmorrill3: Every one of these dots is an actual crash from the fleet. Real wo...

9. bindureddy (Group Score: 98.9 | Individual: 54.7)

Cluster: 3 tweets | Engagement: 2675 (Avg: 422) | Type: Tech

🚨 OPEN SOURCE AI IS LITERALLY UNSTOPPABLE 🚨

The legendary founder of Redis (Antirez) just dropped ds4 - a custom native inference engine built specifically for DeepSeek v4 Flash

This is earth shattering! Here is why:

DeepSeek v4 Flash is a quasi-frontier model with a massive 1M context window

You can now run it LOCALLY on a 128GB Mac using specialized 2-bit quantization

The architecture is reimagined—he moved the KV cache from RAM directly to the SSD disk! 🤯

We already know DeepSeek v4 Flash is insanely good for agentic loops - Now you don't even need the cloud to run it

Closed-source labs are burning tens of billions on massive GPU clusters while single brilliant developers are running frontier-level AI on laptops!

They told us open-source would be worthless against trillion-dollar monopolies

Instead, pure hacker culture + incredible open-weight models are completely rewriting the rules

Open Source will ALWAYS win 💕

See 2 related tweets

@Saboo_Shubham_: OPEN SOURCE AI is killing it.

DeepSeek v4 Flash is a quasi-frontier model with a massive 1M context...

@unwind_ai_: RT @Saboo_Shubham_: OPEN SOURCE AI is killing it.

DeepSeek v4 Flash is a quasi-frontier model with ...

10. sarahmsachs (Group Score: 98.6 | Individual: 35.6)

Cluster: 3 tweets | Engagement: 48 (Avg: 191) | Type: Tech

Underrated truth in applied AI is the winners aren't the ones with the best base model, they're the ones with the tightest data flywheel and the discipline to take action from it. Excited to see what @baseten is bringing customers in this space. there's a lot of fruit for agent labs here.\n\nQT @baseten: Open-source RL libraries break at frontier scale. We built Baseten Loops to fix this.

Loops is a training SDK that takes you from your first RL run to production inference on a single platform:

→ Async RL so training and sampling overlap → 131K+ sequence length for agentic and long-horizon workflows → One command to promote your model to prod → Dedicated infra for predictable, repeatable performance

We're excited to work with @harvey and @EvidenceOpen as early partners.

Early access is open today: https://t.co/cPmBYqCZAa

See 2 related tweets

@dejavucoder: looks like a similar offering to prime intellect hosted training\n\nQT @baseten: Open-source RL libr...
@saranormous: RT @baseten: Open-source RL libraries break at frontier scale. We built Baseten Loops to fix this.

...

11. sairahul1 (Group Score: 98.5 | Individual: 29.0)

Cluster: 4 tweets | Engagement: 13 (Avg: 19) | Type: Tech

What Claude Code features actually make you 10x faster? [Ranked by leverage] From beginner to elite ...

D (Good) • Session resume • IDE integration • @file tagging • --max-turns

C (Powerful) • Hooks automation • Custom slash commands • JSON output mode • MCP integrations

B (Game-changing) ↓↓

I wrote a full guide covering all 35 Claude Code commands, tricks & workflows every developer should know.\n\nQT @sairahul1: https://t.co/3laPA864b3

See 3 related tweets

@sairahul1: 35 CLAUDE CODE COMMANDS & WORKFLOWS EVERY DEVELOPER SHOULD KNOW

claude --dangerously-skip-permis...

@sairahul1: RT @sairahul1: 35 CLAUDE CODE COMMANDS & WORKFLOWS EVERY DEVELOPER SHOULD KNOW

claude --dangerou...

@sairahul1: RT @sairahul1: What Claude Code features actually make you 10x faster? [Ranked by leverage] From beg...

12. Vtrivedy10 (Group Score: 97.5 | Individual: 35.8)

Cluster: 3 tweets | Engagement: 33 (Avg: 31) | Type: Tech

my fave point from here: the earlier you think about your agent as a system that can be measured & improved, the faster you can get a robust agent into production

This isn’t just a technical thing, it’s a human & team thing. Teams that succeed here ask questions like:

“what do I need my agent to do to make our customers happy?”
“What scenarios will my agent encounter in the wild and how can we recreate that in our testing”

Evals are the substrate that determines what your agent does in production. They’re training data for agents because we literally fit our agent to pass Evals via hill-climbing algorithms and human edits to pass failure modes

Once you get your agent into users hands, the eval generation loop compounds. Production data uncovers more issues, these issues turn into Evals, and the agent fits to improve over more cases that could not be captured without real user data.

In the early stages, teams dogfooding their product becomes the feedback signal

Curating evals and running experiments on different agent variants is a muscle every team develops, our goal is to create tooling so every team can create the best agents for their tasks using this foundation\n\nQT @hwchase17: https://t.co/XBOyzAS3Tk

See 2 related tweets

@hwchase17: viv always says it better than me

lots of talk recently of thinking of agents as systems to measure...

@hwchase17: RT @Vtrivedy10: my fave point from here: the earlier you think about your agent as a system that can...

13. omarsar0 (Group Score: 97.5 | Individual: 36.8)

Cluster: 4 tweets | Engagement: 259 (Avg: 95) | Type: Tech

My favourite new stack: Agents + MCP + Markdown + HTML

“Files over apps” is a vibe!\n\nQT @omarsar0: LLM Wikis + HTML Artifacts are insanely powerful.

You should seriously consider this in your workflows.

LLM Wikis captures all the important information that lets you and your agents do meaningful work.

HTML artifacts present that information in interesting ways that allow you to take important actions along with your agents.

My HTML artifacts sit on top of my LLM wikis. They are dynamic and are easily extended as needs arise.

I have hooked my Artifacts to talk to my agents, and similarly, the agents can talk to artifacts.

This has allowed me to build powerful artifacts that reduce my inbox to zero, keep me updated on any topic of interest, fast prototyping, do deep research, design/trigger new experiments, generate figures to improve understanding, schedule research, search relevant information, discover topics, and so much more.

What you see in the clip is not a website. It's a simple interactive HTML artifact.

HTML artifacts are useful for designers, engineers, researchers, students, and anyone working with agents.

Lastly, HTML doesn't replace Markdown. They are a much better combination working together.

See 3 related tweets

@omarsar0: More important takeaway: use both Markdown and HTML.

Your agents will thank you for it.\n\nQT @omar...

@omarsar0: RT @omarsar0: LLM Wikis + HTML Artifacts are insanely powerful.

You should seriously consider this ...

@dair_ai: RT @omarsar0: My favourite new stack: Agents + MCP + Markdown + HTML

“Files over apps” is a vibe!...

14. cgtwts (Group Score: 91.4 | Individual: 34.1)

Cluster: 3 tweets | Engagement: 105 (Avg: 317) | Type: Tech

Jensen Huang, CEO of Nvidia:

“people are teaching their agents to fully run a business and make money”

we’re entering a world where agents handle support, sales, operations, coding and marketing.

the agent takeover is here. entire companies are being compressed into workflows. https://t.co/A1Ch3v12ws\n\nQT @eng_khairallah1: https://t.co/naBnj0ITi1

See 2 related tweets

@RoundtableSpace: Nvidia CEO Jensen Huang : "people are teaching their agents to fully run a business and make money" ...
@eng_khairallah1: RT @cgtwts: Jensen Huang, CEO of Nvidia:

“people are teaching their agents to fully run a business ...

15. Teknium (Group Score: 90.6 | Individual: 64.0)

Cluster: 2 tweets | Engagement: 1944 (Avg: 130) | Type: Tech

We just hit number one globally across all AI apps on OpenRouter.

Super grateful to the nearly 1000 contributors who've helped make Hermes Agent great, thank you!

What do you want to see next?\n\nQT @NousResearch: Hermes Agent is now #1 on the Global @OpenRouter token rankings.

While our journey together has just begun, we'd like to take this opportunity to thank our contributors, supporters, and users for all they have done to get us this far. https://t.co/kA4hPJHKNM

See 1 related tweets

@NousResearch: RT @Teknium: We just hit number one globally across all AI apps on OpenRouter.

Super grateful to th...

16. romainhuet (Group Score: 87.1 | Individual: 29.7)

Cluster: 4 tweets | Engagement: 460 (Avg: 391) | Type: Tech

I really love building iPhone apps in Codex.

Codex can design the screens, write the Swift code with GPT-5.5, run the app in Simulator without opening Xcode, and even click around with computer use to test it!\n\nQT @OmarShahine: Codex is phenomenal at creating Swift iOS apps. I just single shotted an app using goals and got 95% of the way there. Iterating is super fast too. Impressed. Much better than Claude Code.

See 3 related tweets

@Dimillian: I also love building iPhone app with Codex too\n\nQT @romainhuet: I really love building iPhone apps...
@PaulSolt: RT @romainhuet: I really love building iPhone apps in Codex.

Codex can design the screens, write th...

@PaulSolt: RT @OmarShahine: Codex is phenomenal at creating Swift iOS apps. I just single shotted an app using ...

17. latkins (Group Score: 78.6 | Individual: 29.7)

Cluster: 3 tweets | Engagement: 73 (Avg: 104) | Type: Tech

I’ve been consistently impressed by zephyra, and have always felt a kinship with their cause. Beautiful work across the board, and what a slate of releases this week.

Western open weights is going to have a hell of a year.\n\nQT @ZyphraAI: Today we're releasing ZAYA1-VL-8B, our first vision-language model.

ZAYA1-VL-8B is a 700M active / 8B total MoE built on our ZAYA1-8B base trained on @AMD. We achieve strong performance for our size resulting in leading intelligence density and inference efficiency. https://t.co/31BY6rtKvG

See 2 related tweets

@teortaxesTex: They keep going!\n\nQT @ZyphraAI: Today we're releasing ZAYA1-VL-8B, our first vision-language model...
@andersonbcdefg: RT @ZyphraAI: Today we're releasing ZAYA1-VL-8B, our first vision-language model.

ZAYA1-VL-8B is a ...

18. chrysb (Group Score: 72.5 | Individual: 40.6)

Cluster: 2 tweets | Engagement: 141 (Avg: 32) | Type: Tech

i spoke to a founder yesterday - their CTO finally read their agent-made codebase after months and panicked when he realized it was impossible to understand wtf was going on

my rule of thumb is: if your codebase starts written by agents, don’t try to understand it

instead, align at the architectural level before any building happens, and ask the agent to maintain a living architecture diagram of how the system works

there are three altitudes that matter:

Top-level: architecture
Mid-level: patterns & abstractions
Low-level: file-level code

in today’s world, a CTO should be deeply concerned with #1. #2 matters too, but not as critical as #1.

if #1 and #2 are dialed in, #3 is where most of the high leverage agentic gains live.

as long as you understand the architecture and critical interfaces, it becomes much easier to reason about ground truth and meaningfully iterate

understanding and informing the architecture / patterns / abstractions give your codebase maximum longevity and agent maintainability

See 1 related tweets

@garrytan: The fun trick is to have your clankers make diagrams in ASCII of everything and just ask questions u...

19. aakashgupta (Group Score: 68.9 | Individual: 39.3)

Cluster: 2 tweets | Engagement: 118 (Avg: 73) | Type: Tech

The math nobody runs: every fifth hour on a product team is spent rediscovering a decision the team already made.

10 context questions per person per day. 10 minutes lost to each one between the Slack ping, the wait, the answer, and the context switch back. That's 8+ hours per person per week, 20% of every working hour.

Scale to a 50-person org. 400 hours a week. The equivalent of one engineer's full year burned every two months answering "why did we choose X over Y" from Slack threads that scrolled away three weeks ago.

The same gap shows up two other ways. 47% of companies call institutional knowledge loss their top offboarding problem. New hires take 6 to 7 months to feel settled.

Three numbers measuring one thing. The team's reasoning is unsearchable.

The 3-layer architecture in the chart closes that gap. Shared context, shared queries, shared discipline. The same scaffolding that makes an AI agent useful from day one is what makes a new hire useful from day one.

Companies building this stop paying the 20% tax. The ones that don't keep paying it whether they see it or not.\n\nQT @aakashgupta: Hannah Stulberg, a PM at DoorDash, built a shared repo where her team checks in every customer call summary, decision log, and analytics query.

Last week a new engineer needed context on a customer decision from three months ago.

Instead of pinging Hannah and waiting, the engineer opened the repo, asked in natural language, and got the full reasoning in 15 seconds.

Hannah wasn't involved. She wasn't even online.

Every PM book tells you to make yourself indispensable. Hannah did the opposite. She freed herself from being the bottleneck and the team treated her as more valuable.

OpenAI made the same point in their February harness engineering post. That Slack discussion where your team aligned on an architectural pattern? If it isn't discoverable to the agent, it's illegible the same way it would be to a new hire joining three months later.

The numbers back it up. New hires take 6 to 7 months to feel settled. 47% of companies call institutional knowledge loss their top offboarding challenge. 10 context questions a day at 10 minutes each is 8+ hours of productive time gone every week.

I spent the last week studying four implementations: Hannah at DoorDash, Dave Killeen at Pendo, Gabor Meyer at Google, and Carl Vellotti building solo.

Four people, four companies, four different levels of complexity. They all converged on the same three-layer architecture.

Full guide is up with 6 downloadables, including a one-command skill that converts your personal PM OS into a team OS without leaking your personal context.

A personal OS compounds for you. A team OS compounds for everyone.

https://t.co/SPMFsHOjy8

See 1 related tweets

@aakashgupta: Hannah Stulberg, a PM at DoorDash, built a shared repo where her team checks in every customer call ...

20. VKazulkin (Group Score: 68.2 | Individual: 31.8)

Cluster: 3 tweets | Engagement: 712 (Avg: 180) | Type: Tech

RT @sukh_saroy: Anthropic showed a 24-minute workshop on how to actually prompt Claude.

Taught by the people who built it.

Free. No signup. No paywall.

I've watched $300 courses that don't cover what they teach in the first 8 minutes. https://t.co/dHjJTEhmby

See 2 related tweets

@heyshrutimishra: RT @heyshrutimishra: ANTHROPIC JUST PROVED MOST PEOPLE HAVE NO IDEA HOW TO PROMPT CLAUDE. (bookmark)...
@dr_cintas: RT @dr_cintas: Every Claude conversation starts from scratch.

Skills fix that.

And Anthropic just...