Markus N Rabe (@MarkusNRabe) header
Markus N Rabe (@MarkusNRabe) profile picture

Token plumber. My posts do not represent my employer's position.

San Francisco, CA
Joined September 2016
Photo by Markus N Rabe (@MarkusNRabe)
Markus N Rabe retweeted
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must…
10,540
16,296
11,365
87,522
77,142,349
Turns out math was never hard. It was just hard for us.
1
8
449
Markus N Rabe retweeted
With UMA playground, the crazy things Feynman could only ask us to imagine become something you can see. Heat, break, and build at aidemos.atmeta.com/uma
6
8
4
34
6,060
Markus N Rabe retweeted
Today, we’re launching Reve 2.0, the best 4K image model in the world. We invented a new way to generate and edit any image using precise layouts. For the first time, it’s possible to create images you can touch.
266
469
200
5,112
11,954,949
Well, what a week to join Anthropic. I'm impressed by the integrity of this company and proud to become a part of it.
10
910
Hot take: most tokenization “fixes” are over-engineered. We found a simple one that actually works. - drop-in replacement for token embeddings - zero inference overhead - lowers test loss: +8% compute-equivalent Spelling Bee Embeddings: arxiv.org/abs/2601.18030
1
1
7
409
Grateful to my co-authors Zheren Dong and Judith Clymo. This was a fun one.
1
180
Open question for me: does this keep working at 10B/100B+ scale, or do the gains wash out? Would love to see replications.
132
Big thanks to the authors of this gem: arxiv.org/pdf/2510.04212. Will we get this in the official implementation? @tri_dao
1
5
3,522
Markus N Rabe retweeted
New post! This time, about the current state of Long Context Evaluation. I discuss existing benchmarks, what makes a good long context eval, what's missing from existing ones and introduce a new one - LongCodeEdit :)
14
44
10
506
213,553
I was one of the 16 devs in this study. I wanted to speak on my opinions about the causes and mitigation strategies for dev slowdown. I'll say as a "why listen to you?" hook that I experienced a -38% AI-speedup on my assigned issues. I think transparency helps the community.
We ran a randomized controlled trial to see how much AI coding tools speed up experienced open-source developers. The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn't.
97
459
128
3,900
1,809,339
People outside tech starting to realize how quality software and typing/generating code fast are not correlated
Most of my programmer friends tell me that chatbots are a massive improvement and their work is now far far more efficient. As a consumer I have noticed literally no change whatsoever in the quality or price of chatbot-assisted software. What's going on here?
73
384
43
4,460
334,315
Markus N Rabe retweeted
Augment remote agent, your team of virtual interns: 1. Deeply integrated w/ IDE. 2. Spin up remote agents from EXACTLY your current code state and environment. 3. Apply their patches easily to your current local repo. 4. Steer and interrupt them the SAME way you interact with your in-IDE agent. 5. 1-key SSH takeover at any time.
Introducing Remote Agent - cloud-based dev agents with full-codebase context, deep IDE integration, and full toolchain access. Available in VS Code.
1
5
557
Markus N Rabe retweeted
🆕 Live now: Self Coding Agents @colin__flaherty has been working on SOTA coding agents and currently holds the #1 spot on SWE-Bench Verified! youtube.com/watch?v=Iw_3cRf3… It might sound like science fiction but @augmentcode has gone all the way and made coding agents that **BUILD THEMSELVES** :eliezeremoji: Some really useful hard won lessons here exploring: - What does it look like to supervise AI coding agents at scale? - How far can we push AI coding agents - what tasks are out of scope? - What might unlock when AI handles complexity and humans drive innovation?
2
25
3,139
One of my favorite co-workers of all times, @smloos, has openings in her team at Waymo. Take a look! careers.withwaymo.com/jobs/s…
1
196
🧵We just released the #1 open-source agent on the SWE-bench Verified leaderboard by assembling the best of Claude Sonnet 3.7 and O1. Open-source repo here: github.com/augmentcode/augme… Here's how we achieved 65.4% success rate on the hardest coding benchmark in the industry: 🧠👇
7
60
18
275
51,480
Proud to announce that we've just achieved to be 🥇 on SWE-bench verified. swebench.com/#verified Big props to amazing @colin__flaherty for pushing our agents to the limit and beyond. Check out @augmentcode
1
2
23
821
Most code retrieval benchmarks optimize for synthetic problems. But professional software engineers don't solve artificial puzzles--they navigate large, often messy codebases. Optimizing for synthetic benchmarks means climbing the wrong hill. Models that top artificial leaderboards consistently stumble in real-world applications.
5
8
1
25
3,872
The morning routine.
4
7
3
56
4,736
The most important and underreported work in ML is how to create good evals.
6
267