Jun Zhou

We just shipped a big refactor at Trovio... all about treating AI models like tools in a toolbox, not a one-size-fits-all solution. We replaced our single-provider LLM setup with an abstraction layer using the Vercel's AI SDK. Now we can mix and match models based on the task – just by swapping a string. I've found Claude is great at creative work like weekly email summaries to our creators. And then fast classification like matching social posts to content ideas, GPT 5 nano or Gemini flash works great (and costs way less). The temptation when building AI into your product is to pick one model and use it everywhere. But different tasks have wildly different requirements for latency, cost, and quality. Learn from my mistakes, abstract early. Future you will thank you!

8

You can solve simple mazes like this in 1 minute. Frontier models can do that too...right? While designing a new frontier evaluation, we were surprised to find complete failure on basic 2D mazes. In MultiNet 2.0 Preview: Interactive 2D Mazes, we put Claude Opus 4.8, Kimi K2.6, and Qwen 3.6-27B in 50 2D maze environments with mechanisms such as key-doors, switch-gates, distractors and decoys to evaluate models on exploration, planning, causal reasoning, long-horizon action taking, and error recovery capabilities. Excited to share an early research preview, playable demo, and highlights below! Try your hand at solving mazes and see how you do vs AI This work was done by the Fig research team in collaboration with the MultiNet group Manifold Research

23

5 Comments

Hey data nerds! REQUEST FOR COMMENTS! I wrote this proposal to store database documentation for Agents inside the database itself. Very similar to the `AGENTS.md` convention in code repos, but instead, it's stored in a schema/table(s) `_agents._agents`. Since AGENTS.md tends to do be used as a reference to other files that are meant for agents, I'm proposing a simple schema: ``` CREATE TABLE _agents._agents ( resource_type STRING, // `global` entries ARE MUST READ, can define schema, table, domain, column, ... resource_name STRING, description STRING, full_markdownTEXT ); ``` Full spec here including the proposed "system prompt": https://lnkd.in/gYpABNhN Curious to see if this gets any traction.

115

31 Comments

a16z Cultural Leadership Fund just published something worth reading if you're building a software company right now. The argument is simple and brutal: the comfortable middle is over. Either you accelerate growth by 10+ points through genuinely new AI-native products, or you rebuild for 40%+ true operating margins. No middle lane. What struck me reading it is that most of the companies they're talking about are retrofitting. They built something, it worked, and now they're trying to figure out how to bolt AI on top without breaking what already exists. Strelux doesn't have that problem. We have the rare privilege of building with AI as the foundation, not the feature. Every workflow, every decision layer, every piece of the product we're designing is Claude-native from day one. No legacy to protect. No seat-based pricing model to unwind. No internal politics around which team owns the AI roadmap. Just a small team, a clear problem, and the best AI infrastructure available right now as our default stack. I want Strelux to be a 10x Claude-native company. Not because it's a good talking point for a pitch deck, because it's the only way to build something that actually competes in the world this article is describing. The a16z piece is here: https://lnkd.in/eC2EX-d5. Worth 10 minutes. #growth #valuebuilding #claude #strelux #ainative #claudecode

3

1 Comment

AI will not fix your go-to-market motion. It will only expose what is broken. Every week I talk to teams that think adding AI will solve their pipeline problems. The truth is, it rarely does. It just makes the same problems show up faster. If your segmentation is off, AI will send the wrong message to the wrong people. If your timing is off, it will hit inboxes when no one cares. If your content is off, it will do a hundred versions of the same mistake. I saw this happen with one team that built an automated nurture flow before cleaning their data. The messages went out fast, but they went to the wrong segments. Replies dropped to almost zero. They pulled it back, added human review, and the next round started real conversations again. AI is not the strategy. It is the amplifier. It can make good systems great. It can also make bad systems worse. The best AI I have seen still depends on human judgment. It needs people to decide who to talk to, what to say, and when it matters most. Where has AI actually made your go-to-market better? And where did it backfire?

7

6 Comments

3 failure modes of AI output and how to fix it: - Sounds like AI. The most obvious. Fix it by compiling a list or skill that the LLM must follow. A quick search of "signs of AI writing" will turn up a lot — Wikipedia has a massive curated list. - Too long. Have you noticed that AI responses can be really compelling but way, way too verbose? This isn't a seminar! Just tell it to: "cut your response in half and keep the most salient and impactful points." This works. - Lacks human narrative. Hardest to detect and fix. Output sounds perfectly reasonable and yet...something vital is missing. Two things you can do (do both): 1) Create a story around the work you're doing — why you're doing it, who it's for, and how they benefit. Make it about people, not statistical outcomes. 2) Review the output manually, like an editor. Give your edited copy back to the LLM and ask it to internalize the changes. Repeat this.

1

Coding is still the highest-intensity use case for foundation models. Claude Code is clearly the undisputed leader in coding LLMs. And Codex is still playing catch up. So how will the OpenClaw move by Sam Altman change this? The developer market is capped by how many “real” developers exist. Even if every engineer uses an LLM daily, that’s not population-scale. Vibe coders and hobbyists certainly help the case, but still don't make enough of a dent. Agents change the ceiling because they turn “coding” into an internal primitive. Anthropic and OpenAI are taking opposite approaches, and perhaps rightfully so. Claude (the leader) is banning tools like OpenClaw from using Claude. And Codex (playing catch up) is leaning into tools like OpenClaw and allowing people to use their OpenAI subscription to consume coding tokens. The thing is that most people will never open a repo. But they WILL ask an agent do stuff in english from their phones. And to do this work, there are a bunch of things that happen under the hood: tool calls, glue code, retries, state, sandboxes. This is essentially just software engineering but hidden. So wiring OpenClaw’s agent loops to Codex is the point: the real consumer of coding tokens becomes the agent, which can be funneled to a massive user base.

10

6 Comments

NVIDIA AI Releases Orchestrator-8B: A Reinforcement Learning Trained Controller for Efficient Tool and Model Selection Orchestrator 8B is an 8B parameter controller that learns to route across tools and LLMs instead of solving everything with one frontier model. It formulates multi step tool use as a Markov Decision Process, optimizes a multi objective reward that mixes task success, monetary cost, latency and user preferences, and uses ToolScale synthetic tasks for large scale training. On Humanity’s Last Exam, FRAMES and τ² Bench, Orchestrator 8B outperforms GPT 5 tool baselines while running at about 30 percent of their cost and with around 2.5 times lower latency, mainly because it distributes calls across specialist models, web search, retrieval and code execution in a more cost aware way..... Full analysis: https://lnkd.in/g-3xGdQj Paper: https://lnkd.in/gMdFTMJw Model weights: https://lnkd.in/gntkUy4j Repo: https://lnkd.in/gHeKcKTS Project: https://lnkd.in/gaCSryyP Video analysis: https://lnkd.in/gvqJgmtw NVIDIA NVIDIA AI Hongjin SU Shizhe Diao Mingjie Liu Jiacheng Xu Xin Dong Yonggan Fu Peter Belcak Hanrong Ye Hongxu (Danny) Y. Yi Dong Evelina Bakhturina

73

2 Comments