Jarrod Tran

Super Intelligence @ Tesla. Man City and Patriots Fan. All opinions and thoughts are my own.

[Jarrod Tran @ItsJarrodTran]

Anthropic's new Claude dashboards put the SQL under every chart.

That's the part worth copying. A chart without its query is just a number someone wants you to trust. If your team can't open the query and check the join, it's not ready for a decision meeting.

[Jarrod Tran @ItsJarrodTran]

@devXritesh You list human approval for high-value actions. For refunds, where do you set that line: a fixed dollar amount, or something tied to the customer's history?

[Jarrod Tran @ItsJarrodTran]

Once OpenInspect spread through a company, it ran 30 to 70% of the pull requests, and you credit bringing the agent to wherever the work happens. What kind of work stayed off it at the 30% companies?

[cole murray @colemurray]

In January, I pasted a blog post into a coding agent. Nine months later, OpenInspect is running in production at companies from 5 to 500+ engineers.

It's time to hand it off. https://t.co/Dk2fHoiMdT

[Jarrod Tran @ItsJarrodTran]

@PeakGrizzly Your abstain precision is 0.94, and routing Kodiak's unsure answers to an 8B LLM beat both models alone. In that run, what share of calls ended up going to the 8B?

[Jarrod Tran @ItsJarrodTran]

@neal_lathia Ghodsi's clip says enterprise AI is bottlenecked on context, not intelligence, and your reply puts the company's other conversations with that customer above the agent's own history. What decides which of those conversations the agent gets to see on a case?

[Gergely Orosz @GergelyOrosz]

I'm starting to sense a divide in the tech industry:

- People thriving on autonomy who are reinventing how they work and what tools they use with AI: LOTS of work, but they're enjoying it and getting very efficient

- People used to being mentored / trained / or simply told what to do and how. Frustrated with the lack of guidance from management on what to do with AI; also frustrated by seeing some of their peers thriving; increasingly bitter about AI and the state of the industry, as a whole

Neither group seems to be aware that there's another group, or have empathy for the other group

[Jarrod Tran @ItsJarrodTran]

@cwolferesearch The Base Labs post has the before-RL numbers: at least one correct answer in 16 tries on 7.8% of prompts at 1.7B, 31.2% at 4B, 59.4% at 8B. SFT paid off at 1.7B and hurt at 8B. Where would you set the cutoff for skipping SFT?

[Jarrod Tran @ItsJarrodTran]

@roeiherzig In the DreamFAR thread, an operator teleoperates inside the world model to record demos. A human drives it, but it's logged as robot data. Does that count as the human data you mean?

[Jarrod Tran @ItsJarrodTran]

You list 'the agent only drafts and a human sends' as a fix for inbox agents, and also say human approval falls to approval fatigue at 93% accept. Where's the line between those two?

[Sergii Makarevych @sermakarevich]

29 guardrails people run on agents that read strangers' text and can act, sorted by one test: does it hold when the attacker adapts? Text-inspecting ones fold. Capability-removing ones hold. Tried 7 on a support inbox.

Which do you run? Which did I miss? https://t.co/T3JJNjzu9Y

[Jarrod Tran @ItsJarrodTran]

@RobertTLange @GenReasoning One detail in the Avalon-M1 appendix: the org size counts every agent the CEO hired, including any later fired. So the hire counts include agents that got fired. The paper doesn't say how often that happened.

[Ivana @ivanainai]

Unpopular opinion: the people getting the most out of AI are the ones who could do the work without it.

[Jarrod Tran @ItsJarrodTran]

@xiaoyin_chen66 Meta-Harness only accepts an edit if it improves validation accuracy on the active sub-block, and removing that gate dropped stability to -5.5 on workflow drift. Did you try validating each edit against earlier blocks too, before accepting it?

[Jarrod Tran @ItsJarrodTran]

@SuJinyan6 Does the paper split the human win rates by that yes/no cross-domain score, to see whether proposals scored 1 win more often with the two experts?

[Jarrod Tran @ItsJarrodTran]

Send it @Jason

[@jason @Jason]

Do not quote retweet this or I will be forced to give people $42.0 or $420!

This is not engagement farming, it’s 1am and I’m bored AF and we all deserve a 420 break!

Do not quote retweet or you might get smashed with X Money

[Jarrod Tran @ItsJarrodTran]

@sh_reya When a swarm is tuning code against your metric, what catches it gaming the benchmark instead of actually making the system faster?

[Jarrod Tran @ItsJarrodTran]

@FracSlap When 20 agents are working while you're away, what do you check first when you get back to the screen: test results or a diff read?

[Jarrod Tran @ItsJarrodTran]

@lennysan Simple but has made falling asleep at home or traveling so much easier @lennysan https://t.co/brUYFJYlML

[Jarrod Tran @ItsJarrodTran]

@alan_ritter LaRA needs hidden states, so closed models can't be checked yet. is anyone close to a version that works on API-only models?

[Jarrod Tran @ItsJarrodTran]

@sh_reya a person should own the metric definitions. if the agent writes the semantic layer it then queries against, nobody catches a wrong definition until the numbers are already in a deck

[Jarrod Tran @ItsJarrodTran]

@fuxinli2 if the reward comes straight from predicted geometry, what happens when the model guesses the hidden side of an object wrong?

[Jarrod Tran @ItsJarrodTran]

with the serving stack open, a team running V4.1-Flash in-house can check its own cost per token against deepseek's api price instead of guessing

[DeepSeek @deepseek_ai]

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!

Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.

🔗 https://t.co/JGAik1kgk0

1/4