Tony Liu (@tdliu) profile picture

Partner at @costanoavc. Previously @databricks Love data, coffee, food, wine, and basketball. 🇹🇼🇳🇿🇺🇸

Joined June 2013
Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)Photo by Tony Liu (@tdliu)
Announcing Autoresearch Bench A benchmark for coding agents autonomously tackling research problems While existing coding benchmarks are saturating quickly, the difference between models in autoresearch loops is stark
34
56
13
534
58,483
If you’re serious about robot learning you (unfortunately) need to know about video compression. Camera streams dominate data volumes for most datasets at 90+% even when compressed. Video is more complicated to deal with but the size wins are too big to give up. The unit of compression is a Group Of Pictures (GOP). In the simplest case (what you should use in robotics), GOPs start with a keyframe (I-frame) that is followed by several delta frames (P-frames). Delta frames only need to encode the difference to the previous frame which is where the compression win comes from. That means to decode frame 15 of a 30-frame GOP you need to feed all the preceding frames in the GOP to the decoder to get out that one frame. The GOP controls the tradeoff between random access and compression. Why does this matter for robot learning? Because while training, dataloader performance is dominated by fetching and decoding video. To build a streaming dataloader (you need this for large datasets) it needs to take GOPs into consideration when fetching data for a time step. It’s hard enough to build a dataloader that doesn’t starve your GPUs that most teams forgo flexibility. That means researchers at most of the best funded robotics efforts currently wait around for large export jobs before training can start after each change to the dataset mix or the wrong hyperparameter. This situation obviously won’t last since they all know that experiment cycle times is a key lever to fast progress and the competitive pressure is enormous. If you want to compete in this space you need both flexibility and performance.
10
46
2
517
46,895
We built the fastest coding agent for frontend This video is NOT sped up Comment "ami" for 100k free tokens
1,249
139
38
2,388
447,815
i'm building something that's very FAST and good at frontend tasks it's like Lovable for localhost. you prompt → agent edits files on your machine directly rolling out to beta testers now. DM me for priority
58
14
7
402
53,928
Introducing React Grab: Select elements and edit with Cursor/Claude Code Works in your localhost and in any React app
79
136
31
2,058
527,630
introducing Ami comment on the page → make a change to your app
443
309
154
7,089
1,196,915
Physical AI will be the biggest technology shift in history but the current data stack won't get us there. We raised $17M to build the data stack Physical AI needs!
1/ We just raised $17M to build the multimodal data stack for Physical AI! 🚀 Lead: @PointNineCap With: @CostanoaVC, Sunflower Capital, @seedcamp Angels including: @rauchg, @ericjang11, @olivercameron, @wesmckinn , @dessaigne , @abimbhet Thesis: rerun.io/blog/physical-ai-da…
5
6
41
2,046
Tony Liu retweeted
1/ We just raised $17M to build the multimodal data stack for Physical AI! 🚀 Lead: @PointNineCap With: @CostanoaVC, Sunflower Capital, @seedcamp Angels including: @rauchg, @ericjang11, @olivercameron, @wesmckinn , @dessaigne , @abimbhet Thesis: rerun.io/blog/physical-ai-da…
10
25
13
143
29,571
Big news today: @GentraceAI raised our $8M Series A led by @MatrixVC. We’re celebrating by launching Experiments, the first collaborative testing environment for LLM product development.
26
23
13
99
20,084
Introducing React Scan Monitoring (beta) Detect React components that cause the most lag for your users in production
16
67
2
1,011
77,248
Tony Liu retweeted
Pretty good.
Guys fine-tuned Llama 3.1 8B is completely cracked. Just ran it through our fine-tuning test suite and blows GPT-4o mini out of the water on every task. There has never been an open model this small, this good.
11
24
1
264
62,272
Replying to @OpenAI
@OpenAI's new gpt-4o mini is Extremely Based. 1/60th the cost of gpt-4o, and on our internal @OpenPipeAI benchmarks it's actually (slightly) outperforming gpt-4o on 3/4 tasks. Can't wait to fine-tune this baby and see what it can really do.
2
23
726
We'll make your site fast before your competitors at <1% the cost. Email me: [email protected]
Almost anyone in B2B SaaS could dominate their category by achieving frontend rendering speeds below 150ms. Yet, less than 1% of the industry masters this. There's a significant arbitrage opportunity right in front of you. Capitalize on it before everyone else catches on.
16
13
2
369
85,089
(1/4) LLMOps can be boiled down to three things: 1. Make sure the LLM system does what it is supposed to do: eval 2. Make sure the LLM system doesn't do what it is not supposed to do: red-teaming, guardrails 3. Have continuous visibility into the previous two: observability and analytics into dev/test/prod
2
9
67
5,132
Thrilled to be officially recognized as the strongest model on the AlpacaEval leaderboard. 🙂 tatsu-lab.github.io/alpaca_e…
Super excited to announce our Mixture of Agents model+FT pipeline: beats GPT-4, but 25x cheaper! - Humans prefer MoA outputs vs GPT-4 59% of the time - New SOTA on both Arena-Hard (84.8) and Alpaca Eval (LC 68.4) - Optimized for synthetic data generation for fine-tuning 🧵
5
19
2
97
15,087
Super excited to announce our Mixture of Agents model+FT pipeline: beats GPT-4, but 25x cheaper! - Humans prefer MoA outputs vs GPT-4 59% of the time - New SOTA on both Arena-Hard (84.8) and Alpaca Eval (LC 68.4) - Optimized for synthetic data generation for fine-tuning 🧵
10
55
13
484
91,790
Textual Serve is now live. 🥳 Textual apps run really well in the browser. The experience is better than many terminals! github.com/Textualize/textua…
Coming soon to Textual! Run your apps locally via a high-performance multi-process webserver.
6
11
1
81
8,317
🐍 Attention Python devs who have wrestled with dependency management! 🚨 @wuoulf, the creator of Mamba 🔥, is back with Pixi 📦✨, the package manager that will make you forget your package management woes. 😌 In this live demo (full Outerbounds fireside chat in comment), Wolf showcases Pixi's incredible features: 🌿 Effortlessly install packages globally with a single command 🎨 Enjoy a colorful and user-friendly experience when viewing Pixi toml files with bat 📝 Define and manage dependencies, tasks, and metadata with ease using Pixi TOML ⚡ Resolve multiple environments in parallel, saving you precious time 🔧 Seamlessly handle development and production environments using solve groups 🗣️ Harness the power of AI with Whisper CPP integration, including live transcription and multi-language support But the real magic happens when Wolf removes the Pixi lock file and runs pixi run start. 🪄 Over 10 environments are resolved in parallel, showcasing Pixi's incredible speed. 🚀 No more waiting for ages! 🕑 Wolf also dives into the upcoming Rattler Build, which will simplify building conda packages, making it more accessible for everyone. 🛠️🎉 Ready to bid farewell to your package management headaches? Give Pixi a try! 🌠 👉 Visit pixi.sh to download Pixi and explore the documentation 👩‍💻👨‍💻 Join the Pixi community on Discord to learn, collaborate, and get support #PixiPackageManager #PackageManagement #DataScience
4
28
2
97
14,199
I also heavily used Pixi from @prefix_dev again. This was extremely helpful for downloading the right checkpoint, building the custom cuda kernels, downloading sample data, all with a reproducible environment. One of the best dev tools I've used!!
1
2
4
397