Somil Singh | Inside AI

AI & Tech Founder & Creator @ Inside AI⚡ | Discovering what’s next in AI | Tools, trends & future tech | DM for collabs 📧
[email protected]

[Somil Singh | Inside AI @Skywalkerlyzv]

So, starting on X after much internal debating. Show some love ❤️

About me currently:
AI engineer @ Oracle.
Cancer Early Detection ML @ Strand LS.
Robotics @ Cynlr.
Soulsborne Pro @ PS5.

More below 👇

[Somil Singh | Inside AI @Skywalkerlyzv]

@amitwt Here’s mine: https://t.co/IxVJO2cHV3 !

[Somil Singh | Inside AI @Skywalkerlyzv]

Same noise predictor. Different sampling path.
Part 10 checks DDIM's variance budget and what one noise error does to the next step.
Full derivation + runnable checks below. https://t.co/gzmgXRDlpi

[Somil Singh | Inside AI @Skywalkerlyzv]

The model builds a game player, watches it lose and edits the code.
AAArena tests that loop against archived human programs. Six ladders topped, six still out of reach. https://t.co/t1yx3mIgKk

[Somil Singh | Inside AI @Skywalkerlyzv]

Your fine-tuning examples don't all teach something equally new.
PASS allocates a limited budget around weaker concept support. The results include a tie worth keeping in view. https://t.co/ZGhVUpCzQA

[Somil Singh | Inside AI @Skywalkerlyzv]

The agent noticed the conflict. Did it tell you?
This study tracks the gap between seeing contradictory evidence, using tools and writing the final answer. https://t.co/GLq0rgUjA3

[Somil Singh | Inside AI @Skywalkerlyzv]

Same model. Larger group. A different risk question.
This preprint models why testing a few agents may leave the population effect unmeasured. The assumptions matter. https://t.co/UlWzsSPNFc

[Somil Singh | Inside AI @Skywalkerlyzv]

A warning in your IDE can now open a Copilot fix.
The JetBrains update also changes default models, MCP startup and the oldest supported IDE. Details below. https://t.co/MSJAp7C827

[Somil Singh | Inside AI @Skywalkerlyzv]

@ArtificialAnlys Will the tasks test recovery after a missed grasp too?

[Somil Singh | Inside AI @Skywalkerlyzv]

@MaximeRivest Curating the tasks and examples sounds like the fun rabbit hole here.

[Somil Singh | Inside AI @Skywalkerlyzv]

@chrisbbh Does the VM stick around when the agent finishes, or do you pull the changes back and shut it down?

[Somil Singh | Inside AI @Skywalkerlyzv]

@1a1a11a Having students measure latency and cost on their own agents should make those optimization assignments pretty fun.

[Somil Singh | Inside AI @Skywalkerlyzv]

@duckdb The upfront query cost gives the agent something useful to base that wait decision on.

[Somil Singh | Inside AI @Skywalkerlyzv]

@joshuagunnn That pnpm to bun example is a good one. Keeping the old value in history without recalling it again makes sense.

[Somil Singh | Inside AI @Skywalkerlyzv]

@alexzfunk Was pairing or decoding the button events the fiddly part?

[Somil Singh | Inside AI @Skywalkerlyzv]

@btaylor Who the agent represents needs to stay clear through checkout. Especially when it hands off between companies.

[Somil Singh | Inside AI @Skywalkerlyzv]

@fchollet @Busabase4agent Keeping the sources with those docs would be handy when the next agent picks them up.

[Somil Singh | Inside AI @Skywalkerlyzv]

@wheresryan22 The moving parts are what caught my attention. An engine cycle would be fun to teach with this.

[Somil Singh | Inside AI @Skywalkerlyzv]

@TheHumanoidAI Same room, different camera path. Lining those recordings up is where this challenge gets interesting.

[Somil Singh | Inside AI @Skywalkerlyzv]

@ExaAILabs @3blue1brown For the agent search evals, do you measure whether the results actually let it finish the task?

[Somil Singh | Inside AI @Skywalkerlyzv]

@rasbt Advantage stats and entropy in the same walkthrough. That helps connect the equations to what the run is doing.