Brian

E/acc

[Zain @zainhas]

how inference engines actually work - releasing full talk slides!

i cover everything in the lifetime of a request e2e:
> the inference engine
> kv + prefix caching
> continuous batching
> paged attention
> chunked prefill
> sampling
> agentic loops from inside the engine

[Mao Ning 毛宁 @SpoxCHN_MaoNing]

A historic day in China’s space program!

China’s Long March-10B has successfully completed its maiden flight—and recovered its first stage via a sea-based net. This marks the country’s first-ever controlled rocket recovery. A major leap toward reusable launch capabilities. 🚀🌊🇨🇳

[Molly O’Shea @MollySOShea]

NEW: Inside @PsiQuantum's Silicon Photonic Chipset

*Never-Before-Seen*

With Er-Xuan Ping, SVP, Barium Titanate (BTO) Development

Backdrop: the U.S. government just announced a $2B push into domestic quantum computing manufacturing - but what does that actually fund?

The @CommerceGov Department recently awarded PsiQuantum $100M to accelerate development of BTO, giving a rare look into the underlying manufacturing stack required to scale fault-tolerant quantum computing.

We profiled PsiQuantum last September following its $1B Series E, including how the company produces the world’s highest-performing optical switch - a core component of its silicon photonics platform.

PsiQuantum is currently the only company or institution in the world manufacturing this BTO material for optical switches at 300mm scale.

While the company built this silicon photonics platform for fault-tolerant quantum computing (FTQC)*, the implications may extend far beyond quantum itself.

As AI data centers increasingly shift from copper to optical networking, PsiQuantum’s photonics stack could also become foundational infrastructure for next-generation AI systems.

*Fault-tolerant quantum computing (FTQC) refers to quantum computers that can continue operating accurately even when individual quantum bits (“qubits”) are noisy or error-prone.

Shoutout to @PeteShadbolt

[Jaya Gupta @JayaGup10]

https://t.co/5vDrndcmFn

[Neuralink @neuralink]

Our robot is designed to insert hundreds of ultra-fine, flexible threads with thousands of electrodes within microns of targeted neurons while avoiding vasculature and adapting to real-time brain motion.

[Jonathan Nieto @johannieto]

@ycombinator This video is amazing! 💙

[Shenzhen Channel @sz_mediagroup]

The perfect solution for high-altitude cleaning!
This aerial work robot from Envision Energy (远景能源) combines polishing, rust removal, and maintenance in a single unit, making high-altitude work safe and easy.
Envision Energy is a Shanghai-based company that manufactures complete wind turbines and provides energy management software.

@ShLetsMeet @EnergyEnvision @XRoboHub

[Andrej Karpathy @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.

[Mao Ning 毛宁 @SpoxCHN_MaoNing]

Smart construction at an expressway expansion project in south China’s Guangdong. 🛣 No one in the cab!

Unmanned paver + roller working in perfect sync—millimeter precision, maximum efficiency.

[Mao Ning 毛宁 @SpoxCHN_MaoNing]

The world’s first robot has entered the home—as a domestic helper and true member of the family.

This is the future of living. And it starts now.

[a16z @a16z]

Marc Andreessen:

"Overall, the world, especially the western world, it's just stagnant."

"Every once in a while you have somebody that comes along... I actually have an idea how to make things like fundamentally better."

"Nobody licensed us to do any of this."

"Anybody can start, build a product, start a company."

"It's shocking to me how few people actually give it a shot."

"The fate of the world over the next 1500 years is riding on the people who actually want to give it a shot."

@pmarca with @davidsenra

[Brian @QiBein]

https://t.co/GzLOvtalgo

[Ray Dalio @RayDalio]

https://t.co/tjmbT5ytUN

[a16z @a16z]

Product cycles run 10-15 years—and AI's cycle is still just beginning.

In this conversation with a16z's David George and Jen Kha on the state of markets, they cover:

- Why December 2024 was the inflection point for coding
- Why "there are no dark GPUs"
- How AI revenue is growing faster than 100% year-over-year
- Why OpenAI and Anthropic added nearly half the revenue of the entire public software industry in one year

...and more.

00:00 Introduction
02:25 2025 revenue data: 693% growth and why unicorns are real
04:25 Why AI companies outgrow SaaS while spending less
07:15 Adapt or die: Coding tools, org design, and electricity vs. blood
13:09 ARR per employee and what's behind the efficiency numbers
21:42 What Fortune 500 CEOs say vs. what's actually happening
28:24 CapEx, debt, and the AI infrastructure buildout
41:11 Private markets, power laws, and where value is concentrating

@DavidGeorge83 @jkhamehl

[Cannondale @RideCannondale]

Adventure, Smarter and Smoother on the all-new Topstone Carbon 2 with SmartSense 😎

This brilliant all-around gravel bike lets you ride the rough stuff, while staying on the cranks with the momentum to get there first.

🔗 https://t.co/N8rkNI2AtS

#ridecannondale

[a16z @a16z]

It's time to build.

[kepano @kepano]

I'm starting a set of Claude Skills for Obsidian... so far they're centered around helping Claude Code edit .md, .base, and .canvas files
https://t.co/vQWtoGjoxv

[iPaul @iPaulCanada]

原来成年人的财富密码，250年前就写清楚了。10分钟，带你一口气读懂《国富论》，理解分工、价值与自由。

[indigo @indigox]

https://t.co/Jsm4N8AWaY