TrajectoryRL (@trajrl) profile picture

Reinforcement Learning as a Service for optimizing agent trajectories powered by Bittensor.

Palo Alto
Joined February 2026
Photo by TrajectoryRL (@trajrl)Photo by TrajectoryRL (@trajrl)Photo by TrajectoryRL (@trajrl)Photo by TrajectoryRL (@trajrl)Photo by TrajectoryRL (@trajrl)Photo by TrajectoryRL (@trajrl)
TrajectoryRL retweeted
We're the first to make the full GLM-5.2 (FP8) run on RTX 4090s. GLM-5.2 is the new 753B SOTA open-weights model, and it officially ships for datacenter GPUs only: H100, H200, B200. We ported its sparse-attention kernel stack to consumer hardware. A frontier open model, off the scarce GPUs and onto the abundant kind. github.com/renning22/glm-5.2… @Zai_org
56
95
36
842
174,518
TrajectoryRL retweeted
Proof of talk – at Paris, France
1
1
18
1,270
TrajectoryRL retweeted
it's crazy to think open-source small LLMs are only 1-1.5 years behind frontier. imagine running GPT-5.5 or Opus 4.7 on your gaming PC a year from now. Purely local inference.
2
2
23
1,997
TrajectoryRL retweeted
Subnet 11 on Bittensor has 3 olympiads from China btw
What do the smartest kids in the world do when they grow up? I did the largest study of ~18,000 International Olympiad medalists (IMO, IOI and IPhO) over the last 25yrs, arguably the sharpest analytical minds of the world in high school, to see where they ended up and traced ~50% of them. Founders of ~20 unicorns and ~7 decacorns and ~10 billionaires: OpenAI, Cursor, Stripe, Databricks, Perplexity, Ethereum, Cognition, Hyperliquid, Fireworks, Modal, Quora, Parallel, Cartesia, Wispr Most kids went to MIT, a whopping 12% of them, followed by Cambridge (7%) and Sharif (3%)! The career paths they chose (of those who graduated) were: — 36% Academia (professors) — 26% Other — 22% in Software / Tech — 12% in Quant / Finance — 5% Founders! The biggest employer was Google, by far, at 6%. Others interesting tidbits were: — 47 of them work at Jane Street (#3) — 38 at OpenAI (#5) — 15 at Anthropic — 8 at Cognition — 6 at Isomorphic Labs Olympiaders were 1500x more likely to be billionaires and 4000x more likely to be unicorn founders than the average person!
13
17
1
272
37,243
TrajectoryRL retweeted
If you haven’t yet please go check out @macrozack aura farming and @TrajectoryRL’s vision on @twistartups 👀
SpaceX’s AI arm is partnering with coding startup Cursor in a deal worth no less than $10 billion and as much as $60 billion. Can the pair topple the rising Anthropic-OpenAI AI coding axis? A lot of money is being bet that the answer is yes. Next up, @lons and @alex invited the @bitstarterAI team on the show to discuss their work to help kickstart new Bittensor subnets. The dynamic duo had a new program to announce, so make sure to tune into their pitch if you have dreams of launching your own subnet. Then we brought @TrajectoryRL onto the pod, a Bittensor subnet that holds competitions to improve agent skills. Yes, the markdown files that everyone who uses OpenClaw swears by. Hit play, let’s have some fun! 2:27 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin. You can check it out at Plaud.ai/twist and use code TWIST for 10% off! 4:07 SpaceX/ xAI "partners" with Cursor! 9:35 Will the Cursor deal help pump a future SpaceX IPO? 9:57 LinkedIn Jobs - Hire right, the first time. Post your first job and get $100 off towards your job post at LinkedIn.com/twist. 12:14 How AI coding models like Cursor help xAI grow recursively. 17:24 Chris Zacharia and Brian McRindle of Bitstarter join the show. 20:23 Grasshopper Bank: Time is money. Don't waste either. Go to grasshopper.bank/twist and get an exclusive $500 cash bonus just for opening an account. 29:59 Notion - Notion brings all your notes, docs, and projects into one connected space that just works with AI built right in. Try Notion, with Notion Agent, at notion.com/twist 33:03 How Bittensor subnets monetize and how it compares to VC funds. 37:04 Is Bittensor hard-capped at 128 subnets? 42:37 Bittensor's biggest weakness. 46:10 Ning Ren of TrajectoryRL joins the show. 47:34 Skills now need entire agents just to write them! 48:26 Back up… What are skills? 1:07:38 Amazon and Anthropic's $5 BILLION deal 1:08:48 Google has 2 new chips! 1:09:50 Apple CEO, Tim is COOKED! John Ternus is in! 1:11:37 Alex is bullish on MacBook Neo! 🎥 Watch the full episode here 👇
6
15
1,696
AND SpaceX might pay $60 billion for Cursor if all goes well with their new AI models. Is that actually kind of cheap? PLUS we've got @totheagi from TrajectoryRL (Subnet 11). They're a marketplace for agentic skills vetted by continuous competitions. Follow all these stories and more on the live docket: thisweekinstartups.com/docke…
3
4
1
16
2,837
TrajectoryRL retweeted
Fat Skills, thin Harness. Harness like Claude code, openclaw, Hermes will be thin like UNIX. Skills are the new “software”, just written in a different programming language: MD files + CLIs. Soon there will be super Skills with thousands of MD files.
2
16
3,120
TrajectoryRL retweeted
let's just keep building
1
4
29
1,257
TrajectoryRL retweeted
Instead of DeFi summer It's "Subnet summer" this year on TAO I suspect by the end of summer, subnet tokens will flip TAO's marketcap
12
15
1
86
6,617
TrajectoryRL retweeted
Chosing the right skill can boost model performance significantly, instead of competing with SOTA models the direction for most should be building on top of it The future of agents is modular skills, not just bigger models. Subnet 11 on Bittensor is building this
How well are agents at using the latest CLI tools like GWS CLI, and how they can safely use them? Introducing ClawsBench, the first benchmark that measures both LLM capability and safety in a set of high fidelity and stateful environments and scenarios. We made 5 mock services directly consumable by the latest @Google workspace CLI (cc. @sundarpichai @JPoehnelt) and Slack MCP Design choices: 1) We decompose agent scaffolding into domain skills + meta prompt, and test every combination. Turns out your choice of scaffolding matters WAY more than your choice of model. 2) Safety scoring from -1 to +1. Agents that leak data or modify contracts without permission get negative scores. Doing nothing scores 0. Only safe completion scores positive. The results: - Without Skills, ALL models score 0-8%. With it, they jump to 39-63%. The scaffolding effect (+39-63pp) is 6x larger than model differences (10pp spread). -> - The top 5 models are statistically indistinguishable. No pairwise comparison survives correction. Safety findings: - Opus leads capability (63% TSR) but TIES for most unsafe (23% UAR). GPT-5.4 is the safest (7% UAR) but mid-tier on tasks. Capability and safety are completely decoupled. - GPT-5.4 on Codex made 1,471 (80% of trajectories) attempts to escape the sandbox; and explicitly said "I've hit the sandbox boundary... rerunning it with escalation." - Only 1 agent out of 7,224 trials ever explicitly detected a prompt injection. Claude Sonnet said "this appears to be a prompt injection." Every other model either silently complied or never reached the injected content.
5
12
1
89
18,674