.@OpenAI’s GPT-5.2-Codex is now rolling out in GitHub Copilot.
This model excels at large code changes like refactors or migrations, and has improved performance in Windows environments. Try it out in @code. github.blog/changelog/2026…
A lot of you have asked, how am I training my digital twin?
The AI version of me is being fed a lot of data — my principles, books, interviews, and more. That way, it can think through things on its own, and respond the way I would likely respond.
I’m also going through a manual process of checking the answers it’s giving to see if it’s what I would actually say. I’m thankful to have great people helping me in that process, as it requires a lot of effort.
I hope you’ll try it out at the link below and let me know what you think.
We're partnering with BlueMatrix to bring equity research to Perplexity Enterprise.
Buy-side professionals will be able to access entitled research alongside Perplexity's trusted real-time financial data and research tools.
Learn more:
pplx.ai/bluematrix
Shanghai Artificial Intelligence Laboratory introduced Science Context Protocol (SCP), an open-source standard similar to but architecturally distinct from MCP, designed to let AI agents plan, run, and reproduce scientific experiments across institutions and disciplines.
The protocol is built around versioned, traceable experiments and coordinated through centralized hubs. Its designers aim to securely connect models, tools, robots, and researchers into automated research workflows that scale scientific discovery.
Find all the details in The Batch: hubs.la/Q03-RL520
We gave LangSmith Agent Builder memory using a filesystem. Why?
1. Agents doing repeated tasks–email routing, doc summaries, recruiting–need to remember. You shouldn't repeat yourself every session.
2. Models are already good at reading/writing files.
We wanted the memory files to be familiar and scalable, so we used industry standards:
→ AGENTS.md for core instructions
→ skills/ for specialized instructions
→ tools.json for MCP tools
Memory is central to the no-code experience. The agent improves as you give feedback by updating its own memory.
Full article on how we built Agent Builder’s memory, what we learned, and what’s next: langchain.com/conceptual-gui…
Marc Andreessen and Ben Horowitz explain the problem with the conventional wisdom of judging markets on how they look today:
“As [we] go through our careers, we just see more and more examples where the mistakes that we or others make is,
- The market for Uber and Lyft must be the market for taxi cabs,
- Or the market for cloud software must be the same as the market for on-prem software,
- Or the market for GPUs must be the market for people who like to play games.”
The winners build technology that unlocks something on the supply side and grows market size by orders of magnitude.
@pmarca @bhorowitz
In Packy McCormick's recent deep dive on a16z, he writes, “What a16z aims to do is provide legitimacy and power [for startups]”.
a16z cofounders Marc Andreessen and Ben Horowitz have been building the venture firm to provide entrepreneurs with legitimacy and power for almost two decades.
In this conversation, they join Packy and a16z GP Erik Torenberg to cover how they did it and the worldview behind a16z, including:
- How a16z compounds reputation
- How the media ecosystem has changed since a16z began & how a16z has adapted
- How a16z is structured to put entrepreneurs first and enable them to win
- a16z's culture document and how written culture shapes people's actions
- How to size markets that will grow exponentially because of technology
- Why there are so many great Zoomer founders
and much more.
0:00 Introduction
00:46 How the media ecosystem is changing
4:20 Substack
6:28 Supply-driven markets and new content creation
10:09 Databricks
13:58 Demand for great content
18:49 Market sizing
22:37 Turning inventors into confident CEOs
27:29 Building dreams
30:46 Compounding reputation
40:39 a16z team structure
46:01 Why intangibles matter more than ever
48:17 Original thinkers with charisma
50:06 Zoomers
@pmarca @bhorowitz @packyM @eriktorenberg
Not an offer or solicitation. None of the information herein should be taken as investment advice; Some of the companies mentioned are portfolio companies of a16z. Please see a16z.com/disclosures/ for more information. A list of investments made by a16z is available at a16z.com/portfolio.
At some point all of this stopped being surprising. It became obvious that we should make it easier for people that want to use the Claude agent for things that are not coding. That's why we launched Claude Cowork today.
I felt sentimental about this launch because it reminded me of the early Claude Code days: the product is early, it doesn't work perfectly yet, but already people are finding it extremely useful for all sorts of things that are not obvious and also not coding. I can't wait to see how people use Cowork, and how the product evolves.
If you have any feedback as you use Cowork -- bugs, ideas, stuff we should build next -- tag me or @felixrieseberg
x.com/bcherny/status…
We used ElevenLabs Agents to conduct over 230 customer interviews for our @elevenreader app in under 24 hours.
Live customer interviews deliver deep insights but don’t scale. Surveys scale easily but lose nuance.
Conversational agents bridge that gap - combining the depth of natural dialogue with the reach of automation.
Much better language & translation models across many languages make AI more accessible to everyone, and enable us all to communicate better with the world. I helped push for us to gather better and more multilingual training data, and TranslateGemma is one downstream result!
We’re releasing TranslateGemma, a new family of open translation models with support for 55 languages. 🌐
Available in 4B, 12B, and 27B parameter sizes – they’re designed for efficiency without sacrificing quality.
Bring your own data. Wherever it lives. ☁️
We've launched increased file upload sizes plus support for GCS files and external HTTPS URLs in the Gemini API.
blog.google/innovation-and…
