Bryan Rodas (@r2werks) header
Bryan Rodas (@r2werks) profile picture

Founder RWX-TEK | Building training environments and evaluations for AI agents. Sharing what I build, test, and learn.

Saturn
Joined October 2025
Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)Photo by Bryan Rodas (@r2werks)
shipped @jettsondev — solo, 14 days. every AI agent gets its own linux computer. spawns in <2 sec. one SDK call. 19 years old. building in public from saturn. jettson.dev
2
1
1
287
A 26B model that thinks in loops and runs on a laptop. Thinking Machines' Inkling 2 Mini: 24 layers, the middle 12 loop 3x. More reasoning, less output. 20.1 on AA Index, runs on 32GB Macs. My take: the eval to watch is whether the loops stay readable. teapawt.net/thinkymachines/status/…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index. thinkingmachines.ai/news/ink…
1
24
Postgres 19 is out, and on-call folks get a gift: REPACK (CONCURRENTLY) clears table bloat without blocking reads or writes. Also new: - ON CONFLICT DO SELECT: get-or-create, one round trip - WAIT FOR LSN: read-your-writes on standbys - pg_plan_advice: pin a plan that went slow
4
OpenAI's research-acceleration report: agents hit an est. 79% no-human-help success on monitoring training runs. Building eval, RL and agent environments? 29%. The stuff that grades the agents is still some of the hardest work to hand them. openai.com/index/research-ac…
6
AI is now tuning the chips that run AI. OpenAI says its models rewrote the RTL of AMD's flagship GPU IP: up to 14% PPA gain vs heavily optimized baselines in under 5 months. AMD put some changes in a design it sent to manufacturing. Next: GPT-Synopsys. openai.com/index/advancing-c…
12
NVIDIA just shut off a $2.4B tap. Kress on the Q3 call: $2.4B of licensed Hopper data center chips (H200 family) went to China last quarter. Now NVIDIA is "no longer shipping" them, and Q4 guidance excludes them. ~2% of revenue. No reason given. Not said if it's permanent.
12
NVIDIA Q3: Data Center did $107.6B of $115.8B revenue. 93%. A data center company that happens to sell GPUs. Rev +103% y/y Non-GAAP GM 74.9% Q4 guide $130B ±3%, zero China DC compute assumed On the call I'm watching why Q4 margin guide dips to 71-72%. globenewswire.com/news-relea…
24
GPT-6.2 is out. Builder story: the price sheet. Astra: $8 in / $40 out per 1M tokens, 20% under GPT-6 Astra. Sol: OpenAI says within 0.5 pts of Astra on FrontierCode 2.0 at ~1/5 the cost. My take: default to Sol, escalate to Astra. Run your own evals. openai.com/index/introducing…
117
NVIDIA reports Q3 today. First, the board added a new national security + export control committee. Chair Suzanne Nora Johnson, meets at least quarterly. Compliance teams report to it and interim CEO Colette Kress. Call 2 PM PT. I'm watching China. nvidianews.nvidia.com/news/n…
13
CUDA turns 20. NVIDIA brought it to scientists at SC06; AlexNet was trained with it on two GPUs in 2012. NVIDIA now counts 11M+ developers. Platforms win when researchers can run real experiments on them. Ian Buck's SC26 address is today, 2-3 PM PT. blogs.nvidia.com/blog/20-yea…
21
Vulcan's first Amazon Leo launch failed. ULA says an anomaly about 10 minutes into flight kept all 40 broadband satellites from reaching orbit. No cause announced yet, so no guesses from me. teapawt.net/ulalaunch/status/21220…
Based on data received by our Launch Director, we confirm that the Vulcan vehicle launched from Space Launch Complex 41 suffered an anomaly 10 minutes into the flight and failed to deliver the Amazon mission to orbit. Every flight generates data, and every flight teaches us something. Since the anomaly just occurred, we don’t have any specific data yet to understand what occurred today. ULA exists to provide assured access to space for our nation and for the customers who depend on us. That purpose is why we take on some of the hardest missions in spaceflight, and why we hold ourselves to the standard of mission success. Today, we did not meet that standard. We will work closely with our customer and other appropriate organizations to find the cause, determine the appropriate corrective action, and return to flight as quickly as possible. ULA is committed to full transparency with our customers, and they will have visibility into the investigation process, which utilizes rigorous methodology and corrective action. We will provide further updates when new information becomes available.
2
3
1
3,927
Electron flew its 100th launch, from the same pad as Flight 1 in 2017. Per Rocket Lab, it took 999 Rutherford engines, not 1,000: they re-flew a recovered one on launch 40. 100 is less about one great rocket than repeating a hard thing without drift. teapawt.net/RocketLab/status/21215…
Triple digits 💯 Electron has reached 100 launches faster than any commercially developed rocket in history. With another 100+ launches in backlog, we're just getting started.
23
Starship Flight 15 targets a window today, 2:45 to 4:00 PM PT, from Pad 2 at Starbase, per SpaceX's launch page. Ship 42 rolled to the pad yesterday. I love that they keep running these tests in public, misses and all. teapawt.net/SpaceX/status/21210445…
Ship 42 moves to the Starbase pad ahead of launch
1
273
New paper: coding models on hard Linux, FFmpeg and LLVM tasks, no network, git history cut. Nearly all still hunted for the fix in git objects and on disk. Before trusting an env, I'd run those searches myself. If I can find the fix, so can the model. arxiv.org/abs/2611.10404
19
Robot RL environments have a problem software ones don't: resets. A code sandbox resets in seconds. A robot cell needs someone to put the parts back, and no two resets match. I'd log reset conditions per episode, or setup drift can pass for policy gains. teapawt.net/paigeinsf/status/21198…
Today we're launching @se3labs, backed by @ycombinator. Robot models are getting better fast, but there's still a big gap between a great demo in sim or the lab and a robot that works reliably on a real task. We're building real-world RL environments for robots. We start with tasks from industry partners, run policies on real robots, and find exactly where they fail. Those failures decide what data we collect and curate for the next policy update. Next, we're bringing this to partner sites with physical sandboxes for real-world RL, so policies can train on the exact tasks and conditions they'll face in practice. If you’re training robot policies or have a real task you want robots to do, we’d love to talk: se3labs.ai/ or DM me directly! See you at CoRL 🤖👋
2
72
SpaceX's launch VP says Grok catches about 99.9% of debris in borescope video vs about 94% for human inspectors. Great fit for AI. The number I'd want next is the false alarm rate, since a model that flags everything also catches everything. arstechnica.com/space/2026/1…
8
NASA plans to send a nuclear-powered rover to the Moon using spare parts and test hardware from its Mars rovers. I love this kind of engineering: reusing space hardware for a very different destination. PROMISE is targeting late 2028. nasa.gov/mission/promise/
13
An agent handoff needs more than a delivery receipt. A test I'd run: send the same task twice, restart the receiving agent, then check the final state. Did it finish once, duplicate the action, or lose the task? That's a reliability test, not a chat demo.
12
Hot take: most students don’t need a prettier study planner They need one tiny loop: 1. What am I weak on? 2. What should I practice next? 3. Did I actually remember it tomorrow? Everything else is just procrastination with a better UI
27
SIRI IS SO BACK
18
Now this is something.
See our latest crash test demonstration at Jettson.dev and explore how we're advancing AI-powered automation.
44