Marq

Evals, RL environments & AI data | Building benchmarks for binary reasoning.

[Sergio Paniego @SergioPaniego]

how do you post-train an agent?

we laid it out as a 5-step curriculum, from SFT to RL through a real harness, with livestreams, code and docs for each step

https://t.co/qTFRnkkYA7

[Marq @dev_null321]

I can do one better I was rejected from the CIA after going through their test. Cyber security position 😵‍💫

[Conor @_conor_one_n]

I just got rejected in the final round for AI Researcher at @Anthropic.

I applied with 4 months of experience. Here are my takeaways 🧵

[Google DeepMind @GoogleDeepMind]

Gemini 4 Carbon is our next frontier model.

It’s state-of-the-art across cybersecurity and long-horizon capabilities.

[Marq @dev_null321]

@eriknewsha Be careful what you wish for lol

[Chubby♨️ @kimmonismus]

A few thoughts on Anthropic’s approach:

1. Drug discovery will be one of the most significant revenue streams in the future; the goal here is to set the right course early on. Most diseases will become treatable, a market encompassing 7 billion people.

2. The timing is clearly intentional, announced just before the IPO to signal to investors that the business will continue to thrive in the future.

3. This will likely make it clear, even to those outside the "tech nerd" sphere, why AI is so important. It goes beyond just writing code; it solves problems for humanity.

[Marq @dev_null321]

One of the best people to follow

[Daniel Rupawalla @danielrupawalla]

i've found defects in ~65% of the TAU-banking tasks.

these are tasks where the policy isn't enforced correctly. for example, if there are rules the model can violate and not get penalized for (transferring to human reqs, tasks where the tool call can fail and the model is still rewarded, etc). some grading may be directionally correct, but still contains noticeable defects that needs to be hardened prior to use.

high-level defects for people working on benchmarks right now:
> models are capable of taking many out of scope actions without being penalized (only tool-call checks, no db)
> models can directly lie to the customer with no penalty (no communication checks, which for a customer service benchmark feels important, though i didn't label tasks that only fail this as p0)
> many of the tool call checks are naive and if the tool call fails there are no repercussions. i.e. a test of email change may just check if the tool was called, with no field checks or follow through check, so if the tool call fails or if i change the wrong email it may still pass.

to expound a bit more:

(1) i'm going to try and go through every public benchmark. the reason for these audits is that i've become frustrated with people using defective benchmarks to choose what type of data they want to purchase and data co's making flawed data to help a model hill-climb a benchmark.

it is very clear that bad RL data is a contributor to reward hacking, and the benchmarks that inform purchasing decisions lie upstream of that, thus handing a huge power to companies that publish benchmarks.

(2) sierra and the TAU series have had tremendous contributions to tool-calling evals. tau-bench is actually what got me into RL data in the first place more than two years ago. this is in no means to discredit sierra and the effort they've put into creating the benchmark. big inspiration!

(3) please reach out and correct me if i'm wrong on any of these asap! i want to give sierra credit where it is due.

[Marq @dev_null321]

@PatrickToulme sent a dm

[Xplore AI @XploreAi]

Introducing Real World Engineering Bench (RWE-Bench). AI isn’t great at designing rockets, engines, or chips yet. With better evals and RL environments, it will be.

Software was just the beginning. The physical world is next.

[Chubby♨️ @kimmonismus]

Axiom says its AI agents formalized a 15,000-page mathematical proof in just 24 days, largest proofs in mathematics.

Built by over 100 mathematicians across decades, the proof classifies the fundamental building blocks of finite symmetry.

With human guidance, the agents translated it into Lean, making every logical step computer-checkable and repairing errors along the way.

AI tackling an entire interconnected body of mathematics, with results others can verify. A major milestone.

We are witnessing history unfold. Again. Every day now.

[Axiom @axiommathai]

The largest sporadic group, the Monster, has ~8 × 10⁵³ elements.

Related to it is the "Monstrous Moonshine" conjecture, Richard Borcherds’ 1998 Fields Medal work.

Yet for long, there is cloud on the correctness of the proof, as no one person has read the whole - it’s too big.

[Hamel Husain @HamelHusain]

I've seen a lot of pushback to this, with the following complaints:

1. "The benchmark owners already know about the flaws, and are working on it."
2. "Why are they spending time critiquing and not contributing to the benchmark" OR "critiquing known issues are not helpful".
3. "The flawed verdict is too binary"

I find this pushback to be misguided. These benchmarks are used to move millions of dollars of token spend, and many people blindly trust how well these benchmarks generalize to their own products, and don't look any further than the headline numbers.

For example, whenever a new model drops, you'll find thousands of youtube videos analyzing the pareto frontier using one of these benchmarks as the primary quality metric without questioning the benchmark.

I think we would all be better served by inviting skepticism and encouraging people to develop an independent opinion of how well these benchmarks represent the tasks they care about. Concretely:

- For the failing tasks within the benchmark, are they: realistic, have a proper harness, graded properly, sufficiently disambiguated user intent?
- Are models and API calls configured properly (ex: inconsistent context windows, reasoning levels).
- Are the models equipped with a modern harness and minimal retrieval that is sufficiently realistic?
- Does the score even make sense ex: is it just a LLM judge committee that produces a score on a 0-100 scale that isn't aligned with ground truth?

It's often the case you'll see many of these issues while looking at traces of a failing benchmark tasks, and it doesn't take long! Often the **very first trace you look at** has a serious issue.

I also don't think there is an issue with verified vs. flawed verdict. It's good to have opinions. If you look through enough of what's behind their opinions and you find that you usually agree with their verdict, then they have earned your trust, the same way that a benchmark can lose your trust.

I think this eval of benchmarks is a red pill that motivates investing in your own private evals.

[Wenhao Chai @wenhaocha1]

daily reading guys
https://t.co/tDRARBM6Nd

[头雁 @alacheng]

Repo2RLEnv（Hugging Face 开源）和AgentEnv Framework（Scale AI 开源）从不同角度来实现RL环境。

Repo2RLEnv是输入一个 GitHub 仓库，它自动从已合并的 PR、commit 历史中提取任务，生成完整的 Harbor 格式环境（包括沙盒、指令、私有验证器）。验证器直接跑仓库自己的测试用例打分。

AgentEnv Framework（Scale AI 开源）：提供标准化接口定义环境（MCP 工具、沙盒状态）、任务（步骤 DAG）和评分逻辑。可以用插件扩展，把真实应用或模拟工作场景打包成可重复训练的环境。

一些垂直领域可以用AgentEnv Framework这类框架定制RL环境，也会有领域（或是公司）的Repo2RLEnv，根据领域一些特性，工程师做出类似Repo2RLEnv类工具，自动编译出领域的RL环境。这都对传统的数据标注类业务来讲大大的提升了效率，所以硅谷最近出了很多非常小的RL环境团队，大部分都是非常小，但收入很好的。他们就在往这种方向上靠拢。

https://t.co/KV6zD5UAAQ
https://t.co/ooXMqrBnzF

[Qwen AI Fans @QwenAIFans]

💥 You don’t need a 128 GB Mac for a 105 GB Qwen model anymore.

Slotstream streams Qwen3.8-Flash-Next from your SSD — runs on 16–64 GB Apple Silicon Macs. One Swift binary, no Python, offline after download.

Their 48 GB M5 Pro (33 GB budget): 20.91–23.69 tok/s across short / long / coding prompts.

Caveat: Slotstream’s own Oct 5 benches — your Mac/SSD will differ.

via GitHub 👇
https://t.co/eERy6GV6Zq

[Chubby♨️ @kimmonismus]

Anthropic expects AI that rivals top human experts, works autonomously for weeks and controls robots to become widely available within a few years.

Its IPO filing makes super exciting claims:

- Expert-level autonomous systems "within a few years": Anthropic expects AI to match or exceed leading human experts across most professional and scientific disciplines. These systems would work for hours, days or weeks, coordinate with other agents, and operate laboratory equipment and robots through digital interfaces.

- Millions of highly capable agents: Anthropic’s vision is "a country of geniuses in a data center". Vast numbers of agents working in parallel across software, physics, mathematics, biology, medicine and engineering. The company expects exponential progress in AI capabilities to continue.

- AI is already accelerating the development of its successors: Claude helps run experiments, generate and process training data, automate evaluations and safety audits, and write research code. Anthropic calls this "closing the loop": better models help discover improvements that make the next generation more capable. It says over 90% of September’s codebase contributions were written by Claude.

- Scientific progress could accelerate dramatically: Anthropic sees the potential for “compressing decades of scientific progress into years”. It also reports discovering a molecular machine with Claude that might become a new biotechnology tool, potentially for gene editing. That remains an early research finding requiring validation.

- Compute could remain the limiting factor: Anthropic expects increasingly capable models to unlock new categories of work, keeping demand ahead of available compute for the foreseeable future. It also argues that today’s models already have capabilities the wider economy has barely begun to use.

[Xiangyi Li @xdotli]

have we exhausted the mine of low-hanging verifiability for building evals and rl environment?

any thoughts?

[Superman @thesupermannx]

This paper answers why confident idiots run the world.

Cornell University published a psychological paper that explains every terrible boss, every loud influencer, and every broken corporate hierarchy.

It’s called “Unskilled and Unaware of It.”

they tested participants across logic, grammar, and humor to measure the gap between actual performance and self-perception.

And the results are wild.

participants scoring in the bottom 12th percentile consistently believed they were operating in the 62nd percentile.

they were not lying. they were literally incapable of recognizing their own incompetence.

the paper revealed a brutal paradox:

the exact skills required to master a domain are the exact same skills required to accurately evaluate your own competence in that domain.

if you lack the skill, you inherently lack the cognitive capacity to realize you are failing.

this creates a dangerous systemic loop.

the incompetent are blessed with absolute, blinding certainty. their brain protects them from their own ignorance, allowing them to speak first, demand authority, and never second-guess their decisions.

meanwhile, the actual experts suffer the exact inverse.

top performers in the study consistently underestimated their rank. because a task was easy for them, they falsely assumed it was easy for everyone else.

they hesitate. they obsess over nuance. they assume they have nothing unique to add.

we are biologically wired to confuse confidence with competence.

the loudest person in the room gets promoted, gets funded, and gets the microphone, not because they know what they are doing, but because they are physically incapable of seeing what they don't.

the only defense mechanism is building systems that stop rewarding absolute certainty and start demanding proof of work.

[Andrew Curran @AndrewCurran_]

From Jill Lepore's interview with Vanity Fair, in which she argues that Sam Altman is more powerful than Donald Trump, and that The Artificial State controls the United States, and the world.

[Micro Center @microcenter]

🚨Austin Texas🚨 The Grand Opening of your local Micro Center is just weeks away on November 13th! Come celebrate Tech Heaven with food trucks, a DJ, and of course deals. Learn more: https://t.co/ulAlboasE4

[Kimi.ai @Kimi_Moonshot]

Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.

🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.

Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on https://t.co/85jA71X9gw, Kimi Work, Kimi Code, and the Kimi API.

Open weights by Nov 4, 2026.

🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/5iPP0MBej6