Saksham Jain

Applied Scientist Intern @Amazon · Ex-Research Intern @MIT_CSAIL, AIISC

[Saksham Jain @sakshjn]

A few months into my Applied Scientist internship at 𝗔𝗺𝗮𝘇𝗼𝗻 now, so figured i'd share how i got in.

Applied on the portal last june and heard nothing. then a recruiter emailed me in march asking me to actually apply. heres how the interview loop went👇

[Saksham Jain @sakshjn]

@JessicaHullman yeah direction drift is the sneaky one. each agent suggestion looks locally fine so you never feel the goal moving until you're weeks off

[Saksham Jain @sakshjn]

pacing is the flattering version. you ship efficiency and mid-tier bumps when the frontier jump isn't sitting ready to go.

"we chose restraint" and "we don't have an astra-tier model queued" look identical from the outside.

[Theo - t3.gg @theo]

Hot take: this is what pacing looks like.

None of today's releases were Astra or Fable tier. This is intentional.

The point of "pacing" isn't to stop iteration and improvement. The goal is to prevent the development bigger models from spiraling out of control.

Opus and Sol class models are a great place for our focus to go right now. Lots of opportunity for real wins without as much risk :)

[Saksham Jain @sakshjn]

@lydiahallie wait effort level isn't in the cache key? i'd have bet switching mid-session nukes the whole thing

[Saksham Jain @sakshjn]

everyone's mining it for secret instructions but most of 1.9m chars is just tool schemas

that's the context tax every agent turn pays before it even reads a word from you

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

🚰 SYSTEM PROMPT LEAK 🚰

Here's the full system prompt for Claude Opus-5.5!! The total count of everything extracted, including all tools, comes in at over 1.9M characters! 🤯

Lots to dig into here. Enjoy! 🫡

Link: https://t.co/GzISdXkbYq

gg

[Saksham Jain @sakshjn]

@rasbt 128 is tiny. every swa model i've poked at, the window size mattered way less than how many full-attn layers got interleaved

[Saksham Jain @sakshjn]

a model reading gnarly legacy C and surfacing a bug that survived 14 years of human review is the capability worth caring about here.

way harder to fake than a benchmark row, and basically nobody tests for it.

[Peter Steinberger 🦞 @steipete]

Had ChatGPT sometimes crashing on me after updating to macOS 27 and... Astra found a ~14 year old bug in libuv. https://t.co/JE8GssiRXJ

[Saksham Jain @sakshjn]

@_akhaliq ok vlm-to-control is the transfer i actually want to see land, not just another tabletop sim number

[Saksham Jain @sakshjn]

ten agents is the sideshow. the lean proof is the whole result.

an agent claiming a tighter bound is just noise you still have to go verify. a machine-checked one means for once you don't.

[Vals AI @ValsAI]

We asked ten Claude Opus 5.5 agents to devise a faster shortest-path algorithm and prove it in Lean. Within 15 hours, they produced C-HD: a formally verified improvement over the published bounds.

[Saksham Jain @sakshjn]

a per-token price cut means nothing if the model now burns way more tokens per task.

effective cost per finished job is what you actually pay, and that can rise even as the sticker price falls.

[leo 🐾 @synthwavedd]

As expected, unfortunately the only bad thing about the new Opus is it's a token-gobbling monster. Uses the most tokens per Intelligence Index task of any model benchmarked.

The price cut makes more sense here!

[Saksham Jain @sakshjn]

@srush_nlp does conformal actually hold up once your test data drifts off the calibration set? that's the part that always makes me nervous to ship it

[Saksham Jain @sakshjn]

everyone's staring at the 58 on the index but the price is the real weapon here. matching astra on intelligence was coming regardless.

opus at 4/20 undercuts astra and fable by 2.5x, that's what actually forces everyone else to respond.

[Chubby♨️ @kimmonismus]

Let that sink for a moment:

- Opus 5.5 costs 40% less compared to Opus 5
- performs at Fable 5.1 level
- is 30% faster in output
- and you get a banked reset on top of that.

They chose war with OpenAI. We are in such a wild race!

[Saksham Jain @sakshjn]

@yaringal yeah, and it hits reviewers first, they start skimming the second they clock that cadence in an abstract

[Saksham Jain @sakshjn]

@kirodotdev Opus 5.5 on your platform when?

[Saksham Jain @sakshjn]

@SemiAnalysis_ open weights was mindshare for labs with no serving revenue. once you monetize inference, giving weights away is just donating margin

[Saksham Jain @sakshjn]

the interesting part of "daily workhorse" is how much of it is the build harness now

raw weights get you to competent. the scaffolding is what makes it reliable enough to actually live in all day

[Elon Musk @elonmusk]

Grok 4.7 with our Build harness is a strong daily workhorse

[Saksham Jain @sakshjn]

the cheap fast tier is the real product, the flagship is just the number you screenshot once.

three dropping under astra at once means they've stopped pretending the top model was ever the volume play.

[Ziwen @ziwenxu_]

GPT-6-Sol
GPT-6-Luna
GPT-6-Astra-Minor

All three are incoming everyone.

These are the cheaper faster ones that sit under Astra.

Hang on tight and burn all the usage while you have it.

[Saksham Jain @sakshjn]

@Tim_Dettmers static per-layer bit budgets always fell apart for me, early-layer quant error just compounds down the stack

[Saksham Jain @sakshjn]

@Steve_Yegge tpm fits first cause the job is mostly chasing who's blocked on what. agents do that fine, and a bad nudge undoes way cheaper than a bad merge

[Saksham Jain @sakshjn]

Its just hyped because an "OpenAI Researcher" made it. Seems like a pretty standard model to me

[Shriram Krishnamurthi (primary: Bluesky) @ShriramKMurthi]

My poorly-informed take on Jev, which seems pretty impressive, is that it gives us ITTT (If This Then That) on steroids.

[Saksham Jain @sakshjn]

@yaringal LLMs lack the depth and intuition. writing paper is an art and llms are just bad at it