echen

founder @HelloSurgeAI // raising AGI

[echen @echen]

love this table from @theinformation

[echen @echen]

An example sudo L7 prompt:

“Fix the threads page search, it crashes in dev mode.”

That’s it.

L3s are given 400-word specs from their PMs; L7s get asked “make that faster” and know when to ignore their PM.

Example SWE-marathon prompt (full prompt is 5x longer):

[Surge AI @HelloSurgeAI]

When designing sudo L7, we deliberately made the prompts natural and short.

Just 76 words median, compared with 291-427 words across DeepSWE, FrontierSWE, and SWE-Marathon.

Think about a request like “Can we safely cache this?”

To answer it, an engineer needs to understand the system, identify which data can become stale, and decide whether caching is even the right solution.

All of that is unstated in the real world.

Agents need to learn how to handle the decisions engineers have to make when you don’t get a long list of instructions spelled out.

[echen @echen]

Top two trending benchmarks on @paperswithcode are Chartography and GDP.pdf - both built by our Research & Evals team!

[Niels Rogge @NielsRogge]

One thing I like about @paperswithcodee is how it automatically surfaces trending benchmarks for me

e.g. right now the top benchmark for image understanding is Chartography, which includes 100 professional chart-understanding tasks from @HelloSurgeAI

I had no idea this benchmark existed, but both Anthropic and DeepSeek are reporting evals on it for their new models

Find it here: https://t.co/pxwrAzgIj0

[echen @echen]

Learn more about sudo L7:
https://t.co/pi0Rt4VSLf

[echen @echen]

one of my fav findings from sudo L7:

opus 5.5 and gpt-6.1 sol are basically tied at writing correct code. but opus is much better at the behaviors that get you promoted to staff.

on functional correctness:
opus 82.1% ~ sol 82.4%

on thought partnership (e.g., telling you when there’s a better approach than the one you asked it to build):
opus 76.1% >> sol 53.0%

on communication (explaining its decisions and the tradeoffs it makes, plus any limitations):
opus 66.4% >> sol 49.4%

put differently: imagine promoting an engineer because their unit tests pass. that's basically how we've been evaluating coding agents.

but staff engineers do a lot more than that. they also:
- question bad assumptions
- suggest better approaches
- think about tradeoffs
- and help everyone around them make better decisions.

that’s what we want the best agents to do too, and that's what we built sudo L7 to measure.

[Surge AI @HelloSurgeAI]

Could AI actually solve the Riemann Hypothesis?

That's the question that inspired Riemann-bench. We wanted to go beyond Olympiad math and test whether AI could tackle difficult, research-level problems from top mathematicians.

Six months ago, models scored under 5%. Today, they're above 70% (10,000 agents not allowed!), and Navier-Stokes and UGC have now been solved.

With all the recent breakthroughs, we have an internal betting pool with two predictions: the Riemann Hypothesis will be solved within a year, and P vs NP will be the last Millennium problem to fall.

Great to see Riemann-bench featured in this year's @stateofai Report, and who can even imagine next year's?

[echen @echen]

easy: faster FFT
hard: convincing your PM that logging customer banking creds isn't great debugging strategy

[Surge AI @HelloSurgeAI]

We've added Gemini 4 Argon and Muse Wonder to the sudo L7 leaderboard:

Gemini 4 Argon: 18.9% (#7)
Muse Wonder: 15.6% (tied #8)

Claude Opus 5.5 and Sonnet 5.5 continue to lead at 41.7%.

Today's frontier coding agents can write remarkable code. So can a new PhD. Writing great code and being a great engineer are very different things.

Can an agent write idiomatic Rust? Recognize that logging sensitive banking information or storing passwords in plaintext is a terrible idea, even when nobody explicitly warns against it? Proactively warn when a new feature request introduces security risks?

That's what sudo L7 tests, using real-world engineering tasks in private production codebases. Beyond functional correctness, agents have to exercise architectural judgment, anticipate risks, communicate tradeoffs, act as thought partners (don't you wish someone could partner with your PM?), and recognize when technically correct code still isn't ready to ship.

Coding agents are getting very good at writing code. Getting promoted to L7 is the next frontier.

[Surge AI @HelloSurgeAI]

Real software engineering happens in repositories with years of accumulated history.

Legacy code nobody wants to touch. Historical accidents that became architectural decisions. Dependencies that break in surprising ways.

That's why most sudo L7 tasks are set in privately licensed production codebases from real companies.

The tasks are written by staff and principal engineers, former CTOs, and people who've built production infrastructure at places such as Google, Meta, GitHub, and NASA JPL.

Every task starts from a meaningful failure observed in a frontier model. Each one is graded on eight dimensions, from functional correctness to knowing when to push back on a flawed request.

We built sudo L7 to measure the judgment that separates a capable coder from a staff-level engineer.

[echen @echen]

another Surge benchmark on the Muse Wonder model card 👀

Muse scores 33.2% on GDP.pdf. #3 on our leaderboard, right behind Astra

nice work @Meta!

[Mark Zuckerberg @finkd]

Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.

[Surge AI @HelloSurgeAI]

🇫🇷 Le Chonk joins the sudo L7 leaderboard 🇫🇷

We just added @MistralAI's Mistral Large 4 to sudo L7, our benchmark for staff-level software engineering.

Mistral Large 4 scores 13.3%, ahead of Muse Spark 1.3 (12.2%) and Grok 4.7 (10.6%).

sudo L7 tests whether coding agents can go beyond writing correct code to exercise the judgment of experienced engineers: navigating messy systems, spotting hidden risks, and knowing when technically correct code is still the wrong thing to ship.

Allez, le Chonk ! 🥖

Leaderboard ↓
https://t.co/ouCzU2m8mT

[Surge AI @HelloSurgeAI]

An AI agent found every expiring rent discount in a commercial real estate workbook.

Then it missed three break clauses representing roughly $14 million in forecast rent.

The task required following two different paths through a 15-tab workbook. The model followed one and stopped.

Knowing when you’re done is part of the job.

https://t.co/wfNpWtYPzR

[echen @echen]

Learn more: https://t.co/pi0Rt4WqAN

[echen @echen]

introducing sudo L7, our new benchmark for staff-level software engineering.

coding agents are getting very good at being L3s. give them a well-defined ticket and they can write excellent code.

but a staff engineer is valuable for everything that happens around the code.

what should we build? what already exists? what's going to break in prod? what did the tests miss? should we even be doing this?

sudo L7 has 60 tasks, mostly in private production repos from real companies, written by engineers who've actually owned these systems.

expert rubrics grade everything from functional correctness to engineering craft, architectural judgment, thought partnership, and unnecessary complexity.

the best agents succeed on only ~45% of tasks.

coding is only part of the job. for now, we're keeping them at idiot genius L3.

[echen @echen]

🥖 LGTM, le Chonk 🥖

in their latest Mistral Large 4 release, @MistralAI ran a blind Surge human eval across five frontier models - professional software engineers comparing five frontier models on coding quality. the big guy finished #2, first among open-weight models and behind only Opus 5.

most benchmarks can narrowly measure correctness. but human evals get at the deeper stuff: the professional taste, judgment, and maintainability we look for in the real world as well

the difference between passing a set of unit tests vs. getting a "LGTM - ship it"

bien joué, le Chonk ! congrats to the entire Mistral team 🇫🇷

[Mistral AI @MistralAI]

Meet Mistral Large 4, aka Le Chonk.

• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.

Open weights release end of October.

[echen @echen]

they grow up so fast 😊

#1 Riemann-bench
#1 Chartography
#1 GDP.xlsx

congrats, Argon and @GoogleDeepMind! this one’s going on the fridge.

[echen @echen]

blog post: https://t.co/FWY8HJKATA

[echen @echen]

GDP.pdf just got a little brother: introducing GDP.xlsx

the world runs on spreadsheets. somewhere in Sheet2!D18 there’s a vlookup that changes a $14m decision.

GDP.xlsx has 70 real-world tasks across 12 domains. what’s hard is figuring out how the workbook even works:

→ which tab is stale?
→ what does a red cell mean?
→ which comment explains the exceptions?
→ which chart does Finance actually trust?
→ did you remember to check the termination clauses in Sheet 3?

the best frontier agent scores 38.3%.
(congrats Gemini 4 Argon!)

this one’s for you, Kelly.

[echen @echen]

other models: "the tenant's liability is unlimited"
gpt-6.1: "not according to page 14"

great job, sol

[Surge AI @HelloSurgeAI]

A benchmark shouldn’t just be hard to climb.

Climbing it should make the model stronger in ways that transfer.

So what does GDP.pdf teach, beyond PDFs?

Turns out, quite a lot.

[echen @echen]

Blog post: https://t.co/5hYaWw9ojl

Paper: https://t.co/Gom6DmssFZ

[echen @echen]

New research:
What Does GDP.pdf Teach, Beyond PDFs?

GDP.pdf has become a widely used frontier eval, so we wanted to know: if we train on the underlying skill, do we just get better at PDFs? Or do we learn something more fundamental?

So we post-trained Kimi K2.7 on 1,465 GDP.pdf companion tasks.

As expected, GDP.pdf improved substantially:
11.0% → 24.5%

But GDPval improved too: full-task success more than doubled, and 90% rubric pass rate jumped +13.2pp - including on tasks with no PDFs, in an environment with files, tools, spreadsheets, and finished deliverables.

Why? What behavior transferred?

Before training, Kimi often grabbed the first plausible source and started working.
After training, it looked around first: what's here, what matters, is anything missing?

On one task, the base model opened the first weekly timesheet and prematurely started calculating a monthly report. Incorrect!

The trained model correctly found all four weeks first, combined them, and then started working.

My favorite way to think about this:
A PDF is a small evidence environment. A good professional PDF forces you to figure out which section a fact lives in, which charts matter, how the pieces fit together, and whether you have enough evidence from Sections 1 and 2 to act.

Based on trajectory analysis, our hypothesis is that training on rich PDFs taught the model to gather all the evidence before it started acting.

After 1,465 tasks, Kimi kept that habit even when the task moved beyond PDFs.