Lilu

Building @falconer_ai. Former TL at Meta Superintelligence Labs. Ex-ByteDancer and Googler.

[Quinn @superonbrand]

https://t.co/8GXrrz02S0

[Lilu @xulilu]

@larrylv We work through nights like these so our children and grandchildren may inherit gentler days, with more time to live than to labor.

[Max For AI @MaxForAI]

笑死我了，Nvidia昨天收购Hugging Face居然有这么两个彩蛋

NVIDIA 收购 Hugging Face 的交易的最终价格是 $12,930,300,000。

把前六位 129303 当成 Unicode 码点，就是 U+1F917，对应的 Emoji 正好是：🤗

也就是 Hugging Face的logo。

Hugging Face 联合创始人 Thomas Wolf 在官宣时就提示过，收购价格里同时藏了 Hugging Face 和 NVIDIA 两个彩蛋。

第二层也很直接：把 129303 写成十六进制颜色 #129303，得到的是一块很接近 NVIDIA 标志色的绿色。

一笔接近 130 亿美元的收购，估值最后几位都没浪费。

这就是技术的浪漫啊

[Lilu @xulilu]

A reset is like a free refill. If your cup (plan) is smaller than your homie’s, every refill gives you less too.

Extra resets are a perk, but somehow make the cheaper plan feel like you’re missing out.

Clever move by OpenAI.

[Lilu @xulilu]

OpenAI's "reset" trick worked insanely well for Codex growth.

Apr 7 → Jul 13: +4M users in ~97 days
Aug 21 → Aug 31: +5M users in just ~10 days

And August is when OpenAI really started leaning hard into the reset trick.

https://t.co/L2k6xbXqu4

[Lilu @xulilu]

Hi @Caltrain, for more than a month, the Wi-Fi has consistently gone down when northbound trains approach Menlo Park. Not sure if you’re aware of the issue, but it would be great if someone could look into it.

[Dave Nunez @capitaltruist]

I hate meetings and I hate DMs. Work with me if you never want to talk to me

[Quinn @superonbrand]

Does my manager hate me? Just realized we DM'd for the first time in months. He's the founder and I run marketing...should I be worried?

He hates interruptions and truly means it. We get along great, but tbh I don't hate the lack of DMs...

For the past few months, we've only communicated in a channel with our company's knowledge agent @falconer_ai. It's saved us hours of interruptions and context switching.

"@ falcon is this product claim still true? Check the code and docs."

"@ falcon what did we promise the creative team, and what still needs a decision?"

"@ falcon turn this thread into a creative brief, landing page, and launch plan before we lose the idea."

"@ falcon find the latest doc and update it with what we decided"

When Falconer handles the questions, we go into our 1:1s already on the same page.

@capitaltruist and I get to spend our 1:1s catching up as real people, thinking about the next thing at Falconer, and waiting in the—notoriously long—Golden Goat line.

[Dave Nunez @capitaltruist]

"Ruined the em dash" will be on my tombstone.

A few years ago I constantly complained to the labs about lack of proper em dash usage. They really pounced on that feedback.

[Lilu @xulilu]

@tobi People are sharing workarounds, but that misses the point. Engineers contributing to the same repo should be free to choose their own tools. I want a healthy coding agent ecosystem, not one company using its lead to push out competitors.

[tobi lutke @tobi]

I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc.

Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.

[Percy Liang @percyliang]

🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.

[Dave Nunez @capitaltruist]

In a few years, every big software company will be competing with the major labs.

If you give the labs your data, your context, your business intelligence (whether it's explicit or tacit), you're giving up on your long-term prospects. After true ASI, your half-life is shot.

RSI is coming, and it will eat your business if exposed to it. But it's a paper tiger if it can only feast on synthetic data and purchased slack threads from dead startups.

@harvey, @tryramp, @cursor were years ahead of everyone else with AI. They're showing what the next few years will look like for successful companies by owning your own context and intelligence.

[Giorgio @iiamgio]

My fear with AI coding agents taking over is that our library ecosystem will stagnate.

We’re unconsciously leaning towards popular frameworks that agents know best and experimenting less with niche tools.

The “new JavaScript framework” meme is dead.

[Lilu @xulilu]

@capitaltruist Spotted the Chinese edition of this book in a bookstore in China and bought it immediately. I promise I’ll find time to actually read it.

[Dave Nunez @capitaltruist]

Since writing is a hot topic these days, here are two easy sippers for *anyone* in tech—whether beginner or advanced—who wants to immediately uplevel their writing skills.

[Dave Nunez @capitaltruist]

One benefit of coming up with a cool company name is that people give you cool and unhinged shit related to the name.

[Lilu @xulilu]

@FireFlySquid380 I may not have explained it clearly. It wasn’t a cybersecurity task, just ordinary engineering work that GPT flagged as one. I still need to get it done, so I may have to use another model.

[Lilu @xulilu]

OpenAI: Security professionals can apply for access.

Security rookies: Cool, I’ll take the code and data to Kimi K3.

Open-weight models are already close to GPT-5.6, and rookies still have bills to pay. I can’t tell if this reduces risk or just reroutes it.

[Falconer @falconer_ai]

Give every team and agent a focused home for the knowledge they need.

Spaces reduce noise while keeping shared work connected and searchable.

- Agents only see the context each user can access
- Set permissions once at the space level
- Keep sensitive team knowledge restricted