Henry Mao (@Calclavia) header
Henry Mao (@Calclavia) profile picture

Building agent simulations https://teapawt.org/t.co/GTNJu4Ih6i, backed by @spc; Prev. co-founded https://teapawt.org/t.co/PrvwcJhGOz, https://teapawt.org/t.co/jUkhv9tNKF (acquired)

Joined May 2009
Smithery has been acquired by Arcade! It’s been an incredible year building Smithery alongside the MCP community. I'm proud that we've built something that helped process millions of tool calls and made 10K+ MCPs from developers around the world more discoverable. My co-founder @kamathematic is joining Arcade to keep building toward our vision of making tool calls the new clicks. As for me, I’ll be taking some time to explore what’s next. More soon!
Replying to @SmitheryDotAI
@SmitheryDotAI, the leading public registry and hosting platform for MCP, is now part of Arcade.dev. Smithery made finding and running an MCP server take one click. Pairing that developer experience with the security, reliability, and governance that Arcade is known for, gives enterprises both to lead the next generation of agentic AI. Smithery co-founder @kamathematic is joining us to keep building it. More from @TheMostlyGreat: arcade.dev/blog/smithery-joi…
27
4
87
10,580
Henry Mao retweeted
There's one claim in this essay that feels unconvincing: Dwarkesh suggests that narrow superintelligence will create its own RL environments and simulations of entire digital economies, allowing it to leap from verifiable to unverifiable domains. Having sold RL environments and spoken with people building them, I think there’s a sim-to-real fidelity problem that's often overlooked. Useful synthetic data, in most cases, doesn't come from thin air. Even if you don't hire expert labelers, synthetics are often produced by the RLE engineers who draw from their prior experience or interviews with experts. Signal has to come from somewhere external to the model under training, whether it's encoded via simulations grounded in the real world, traces you can mine, or records written down by people. AI also needs to touch grass! That said, RSI will improve sample efficiency, so we'll need fewer experts and every experience becomes higher leverage.
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
5
2
2
27
4,090
There's one claim in this essay that feels unconvincing: Dwarkesh suggests that narrow superintelligence will create its own RL environments and simulations of entire digital economies, allowing it to leap from verifiable to unverifiable domains. Having sold RL environments and spoken with people building them, I think there’s a sim-to-real fidelity problem that's often overlooked. Useful synthetic data, in most cases, doesn't come from thin air. Even if you don't hire expert labelers, synthetics are often produced by the RLE engineers who draw from their prior experience or interviews with experts. Signal has to come from somewhere external to the model under training, whether it's encoded via simulations grounded in the real world, traces you can mine, or records written down by people. AI also needs to touch grass! That said, RSI will improve sample efficiency, so we'll need fewer experts and every experience becomes higher leverage.
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
5
2
2
27
4,090
Henry Mao retweeted
Tardigrade v0.58.0 comes with a fully Effect native plugin system for agents called libraries: Tardigrade library = Effect service + RPC contract + prompt blocks Application hosts can plug different service implementations using Effect layers. Libraries also let you append static/dynamic prompt blocks into the system or user message (cache aware). For example, the latest state of the current working directory or relevant memory based on user message. This gives strong type level guarantees, first-class dependency injection and integration with the rest of your Effect codebase. All fully durable! Currently it's closely tied to tools, but it would soon be extended to agent events. Read more here: tardigrade.sh/docs/concepts/…
5
4
81
2,591
I tried reading a physical book this week, and suddenly everything felt so slow. Maybe these are the first signs of "AI withdrawal" from managing many parallel agents.
1
5
319
Henry Mao retweeted
Tardigrade v0.54.0 is now on Effect v4 stable release (finally)! We just migrated to v4.1.0 from v4.0.0-rc, and will be tracking future versions closely
3
3
78
3,278
It would be funny if, in the long term, SGD turned out to be just a crutch to bootstrap ourselves to "LLM descent"
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2
9
579
We also redesigned our landing page!
Tardigrade 0.52.0 comes with a batteries included agent built from the core primitives. It's been hardened with a lot of production usage. A cool thing is the concept of libraries, which uses Effect TS to inject service dependencies. Almost like npm for agents. For example, in the snippet below, you can use workspace library and layer it with a bun runtime, a virtual workspace with sqlite and isolates, or a full container with a mounted bucket. Your entire agent is a. typed effect program. All of these are also exported as platform layers. Check out more here: tardigrade.sh/docs/agent
5
1
1
37
11,961
Somehow it's not too different from buying data
this is how finding GPUs goes 1. someone says they have GPUs 2. turns out they are a broker for some other guy 3. that guy doesn't have the GPUs either, they are waiting for a buyer so they can place an order 5. they don't even have the space yet, that is owned by some other guy who's upgrading it one of these guys is always named carl
1
3
559
Henry Mao retweeted
Combining Cloudflare's Clairic and Jev beats OpenAI's privacy filter at detecting private data in legal documents. Hunch is that it applies across domains. Using extraction and decision models together seems like a generally good pattern! Wrote my notes over here: clavia.ai/blog/detecting-pri…
1
4
3
92
13,025
Henry Mao retweeted
our startup was acquired today :) jenni ai is now part of the digital science family we fucking did it!!!!! i could go on and on about why this acquisition is so perfect, but i’ll save that for another time for now, i want to focus on the fact that we built jenni in public from day one all the way to exit there was honestly very little i didn’t share, so i hope this archive of our journey can help a few founders on their own fantastical journeys thank you to everyone who's followed along and witnessed all the pivots, near bankruptcies, and general hijinks that come with running a company despite the roller coaster, we ended by hitting our highest MRR ever shortly before deal close 🥳 it's been a crazy ride, you all helped me live out my dream, truly i owe my success to my team, my investors, and my family for all of their support throughout the years not too bad for a college dropout who had to move back in with his parents!
466
63
28
2,596
130,395
One of the most annoying failure modes frontier models have right now is overcorrection: When you tell it to do something, it forgets the bigger goal, overcorrects on the specific ask, and visibly marks the correction.
1
5
441
If you're building evals and benchmarks, what do you currently use?
7%Braintrust
0%Raindrop
13%Langfuse/Langchain
80%Custom tooling
15 votes • Final results
1
363
Henry Mao retweeted
We’re betting our $5.5M seed round that the next leap in AI coding won’t come from OpenAI or Anthropic. It’ll come from a new feedback loop that learns from you. Introducing Beacon: open-source, self-improving memory for coding agents. Your sessions capture how you debug and solve problems. Today, that context gets lost and you have to re-explain things every time you switch tools. Beacon continuously distills your sessions into reusable skills across Claude Code, Cursor, Codex, and 20+ harnesses, so your next agent can learn from the last. Beacon is free and OSS 🌟 github.com/Asymptote-Labs/ag… Why should the frontier labs be the only ones winning with your sessions? You already did the work. Now let your coding agents learn from you.
329
167
143
1,571
634,192
Traditional ML is cool again!
Today we introduced a novel system one model, Clairic, based on regression instead of classification (i.e. decision models like Clef). This enables estimation of continuous values without having to exhaustively list choices upfront. It also enables very efficient extraction of structured data from unstructured inputs. Open weights coming soon!
1
324
Before agents: I make algorithms After agents: I make recipes
1
314
Henry Mao retweeted
Tardigrade is a durable harness framework purely built on Effect primitives. We solved it by write-ahead logging of effects and using event sourcing to derive state. We also used a bunch of cool Effect features like SchemaAST (for incremental JSON validation) to make it more performant than pi-durable on warm turns as events grow.
Wanted to share a few thoughts regarding the Effect part of this discussion, starting with saying that I do broadly agree with everything that was said and that's exactly why I'd like to add some "maintainer" perspective on the choices.
5
6
2
101
7,475
When you hire a team of meat proxies:
2
11
617
Henry Mao retweeted
/1 Introducing Tardigrade Atoms! ⚛️ Atomic state management for durable cloud agents, powered by @EffectTS_ v4 atoms and event sourcing If you've used @dai_shi's Jotai atoms, this will be a delightful way to manage agent's state and side effects. Everything is an atom! github.com/clavia-labs/tardi…
4
8
118
4,720
It's great that we're seeing more "decision models" aka classifiers being polished
Introducing Mercury Decide: the most intelligent decision model on @OpenRouter (JevBench v1.4) and one of the fastest. Up to 14 decisions per second. Watch it go head-to-head with Jev in chess. Free for early access on OpenRouter: openrouter.ai/inception/merc…
1
1
1,599