Ryan Castner retweeted
It's wild to think about how massive 1M token context windows in LLMs really are
That's roughly equivalent to:
- The complete works of Shakespeare
- 11 hours of audio
- A 5-minute session fixing some TypeScript issue
Ryan Castner retweeted
I'm usually not one to write thought pieces without much technical depth. But here we go.
Slow the fuck down.
mariozechner.at/posts/2026-0…
Ryan Castner retweeted
This is @RepoPrompt 2.0
A fully integrated agent, that makes it seamless to use RP's powerful MCP tools, with a built-in oracle and context builder.
A first class experience showcasing how much better and efficient your agents can be with good context engineering tools.
Ryan Castner retweeted
Ralph Wiggum is the craziest thing to happen in the coding agents space in 2026 so far
not the technique itself. just the fact that someone was like "here's a technique it's called the Ralph Wiggum technique because of the Simpsons guy" and everyone was like "OK sounds good"
Ryan Castner retweeted
This is the AlphaGo moment for software agents.
Every AI coding agent today, from Cursor to Devin to Claude Code, learns from human traces. GitHub issues. Pull requests. Test suites written by developers. Someone had to write all of it first.
This creates a hard ceiling. SWE-bench Verified required 93 professional developers to manually screen 1,699 samples just to produce 500 usable problems. The original SWE-bench needed human-curated issue descriptions, pass-to-pass tests, fail-to-pass tests. The data scales linearly with human effort, and human effort doesn’t scale.
SSR breaks that constraint entirely. One model plays two roles: bug injector and bug solver. It explores real codebases, discovers how tests work on its own, creates progressively harder bugs by removing code or manipulating repository history, then learns to fix what it broke. The injector must produce bugs that fail tests through semantic errors, not syntax mistakes. The solver gets no natural language description, just the raw codebase and a test patch.
No human labeling. No curated issues. No pre-existing test suites. Just Docker images with source code.
The results: +10.4 points over human-data baselines on SWE-bench Verified. +7.8 on SWE-Bench Pro, which contains 731 enterprise-level tasks averaging 107 lines of code across 4+ files. The model trained entirely on self-generated bugs still generalizes to solving natural language issues written by humans that it never saw during training.
They also introduced higher-order bugs constructed from the solver’s own failed repair attempts. When the solver fails, those failure patterns become new training problems. The curriculum gets harder automatically.
This is the same loop that took Go from “decades away” to superhuman in 18 months. AlphaGo didn’t need human games to learn. It generated its own curriculum by playing against itself, discovered strategies no human had conceived, and scaled past every grandmaster on the planet.
We’re watching the same pattern emerge in code. The bottleneck was always data, and the assumption was that useful coding data required human developers to create it. SSR suggests that assumption is wrong. Raw repositories plus self-play can generate unlimited training signal.
Meta trained this on 16 million token batches across 150 steps using H100 clusters. The infrastructure exists to scale this dramatically further. More repositories, more self-play iterations, harder bugs.
GitHub has 200 million repositories. The training data ceiling just disappeared.
Ryan Castner retweeted
A few things I've noticed as all devs write code with AI.
When you write foundational / architectural code of a new project by hand, you "feel" the code pushing back if your abstraction isn't right. You feel when something is harder than it should be. The code is telling you it's not in the right shape. Good engineers are sensitive to this.
When you're using an LLM, you keep pushing right through this in a way that feels like you're making progress, and it may even be directionally correct in a sense, but the underlying foundation of it all is actually bad in a way that either kills progress of the LLM later as it buckles under the complexity it has created or destroys your ability to maintain the code long term.
Related to this, I see a general restlessness with just sitting and thinking about a problem for a while.
As I've been working on a new library here at Laravel, there have been days where it feels like I mainly just stare at my screen thinking about something. When Claude Code is at your fingertips, it's tempting to just start yapping into the terminal and watching code come out the other end. Again, directionally correct in some ways, but often doesn't land on the elegant solution that is waiting to be discovered.
Ryan Castner retweeted
📣 TanStack AI Alpha is here!
✨ Framework agnostic
🤖 Provider agnostic
🧠 Type safe
🔧 Isomorphic tools
🛠 Devtools
🌐 Open protocol
📦 JS, Python, PHP
⚛️ React, Solid, Vanilla
🌀 OpenAI, Anthropic, Gemini, Ollama, ++
Official blog post: 🔗⬇︎🧵
Ryan Castner retweeted
I spent the last 60 days working at Cursor. It's been one of the most thrilling phases of my professional life.
There's a lot of mystique around the company. Over the last two months, some things matched my expectations; many did not.
I wrote an essay for @joincolossus about things that have surprised me about the company and its culture so far.
joincolossus.com/article/ins…
Ryan Castner retweeted
This is @RepoPrompt 1.5
The new Context Builder connects to your agent of choice, using your existing subscriptions, to fully automate the process of building the perfect context for a given token budget.
All of Repo Prompt's power, fully automated, with 1 click
Ryan Castner retweeted
In our investigation, we uncovered three separate bugs. They were partly overlapping, making diagnosis even trickier.
We've now resolved all three bugs and written a technical report on what happened, which you can find here: anthropic.com/engineering/a-…
Ryan Castner retweeted
👀 If you thought Opus was a bit mid lately
Anthropic tried a new inference stack and that caused degraded performance
status.anthropic.com/inciden…
Ryan Castner retweeted
To put it mildly, the past week at Windsurf has been crazy. There have been a lot of different rumors and reports, so I want to share a transparent account of how it actually went down.
Before I start, I just want to say that Varun and Douglas were great founders and this company meant a lot to them, and it should be acknowledged that this whole situation must have been difficult for them as well.
One week ago, last Friday, I walked into the office for our all-hands, where our ~250 people were expecting to hear that we were getting acquired by OpenAI. By that time, I had already learned what was really about to happen and had broken the news to Graham, the new President, and Kevin, the new CTO. You can imagine the shock when the team found out.
It was my job to explain to the company our path forward as a company. In my view, there were a few options: try to raise more money (VCs were offering), try to sell the company (we had interest from multiple parties), try to distribute (although there were liabilities to wind down), or try to keep running the company. Although we had lost some great people and taken a serious blow to morale, we still had all of our IP, product, and strong talent including an excellent GTM machine - the core components of the company were still there.
The mood was very bleak. Some people were upset about financial outcomes or colleagues leaving, while others were worried about the future. A few were in tears, and the Q&A had been understandably hostile. In distress, people asked if we could distribute the cash immediately, but we also needed it to pay bills and keep the product working for customers.
After trying to steady the team as much as we could, Graham and I spent the rest of the evening on calls, trying to identify every available option.
Out of the blue, we got a text and email, from Scott and Russell, with the message, “Chat?” It was around 5:30pm, exactly this time last week.
I immediately called Graham and told him I thought this combination made sense. Cognition had been the one team our people had respected the most. While they had overinvested in engineering, they had frankly underinvested in GTM and Marketing, and our teams in those functions are nothing short of world class. On the other hand, we now were missing a Core Engineering team, and there’s no better group of AI engineers than the lineup Cognition has assembled.
Then, there was the product logic. Devin would benefit from a foreground synchronous agent, while we needed a remote asynchronous agent. The teams and products together would be able to create an unrivaled end to end platform.
We took the Cognition approach very seriously from the start and launched right into negotiations. Scott and his team moved fast, and while the timeline was exploding with memes and commentary between Friday and Monday, Scott and Russell spent that time in our office, working tirelessly to get a deal done in record time. Saturday rolled in, and I brought Kevin to the office. At this point, we were still having 1:1s with our enterprise engineers to retain them, and at the same time, we were trying to get more information to diligence Cognition while reviewing other potential partners.
Saturday, I was still getting inbound interest from potential acquirers, including one we had all looked up to for a long time. But by then, Scott was already in our conference room with physical papers to sign and was handing me a pen. We had already made up our minds that there was no better partner than Cognition, even with other excellent and impressive companies interested. We quickly brought in lawyers to review the LOI and signed that day, a little over 24 hours after Scott’s cold outreach.
Saturday had been spent understanding each other’s businesses, and when the sun came up Sunday morning while we were in the office, we prepared to move on to getting the deal finalized. The Windsurf team had been through too many twists and turns, so both Scott and I felt that the next transition for them had to truly be final.
The other priority during this process, which Scott and I aligned on and was one of the things that helped me realize he was the right partner, was that we needed to take care of all Windsurf employees. Their work and talent had gotten us to this point; they deserved to be paid for that, and we wanted to give them a better price than any of the previous scenarios. That resulted in a key part of the deal: structuring it to give a payout to every employee, to waive all cliffs, and to accelerate all vesting for Windsurf equity.
On Sunday, an army of lawyers from both sides descended on the office, having been challenged by Scott and Russell to get this deal finalized within 24 hours. We ate and slept (or at least tried to grab quick naps between discussions) in the office through the weekend to get it done. Our teams and lawyers stayed up all night Sunday working out the final details. On Monday morning, we went over every detail again, got board approvals, and the documents were ready to sign. One of our lawyers said it was one of the fastest deals they’d seen.
Monday 9:30am, we signed our definitive agreement for Cognition to acquire Windsurf. Scott had given a heads up to Cognition employees, and we had another all-hands scheduled for Windsurf employees at 10am that morning. After the traumatic announcement on Friday, Scott and Russell wanted us to be able to open Monday with good news and we wanted to handle this new announcement in a much better way for employees. In fact Russell took a red eye the night before and somehow made it to the Austin office just in time.
We held Monday’s all-hands in the same room as Friday’s, and this time Scott was by my side at the front of the room. It was a blur, but the highlights I’ll always remember were getting to tell employees what we had negotiated for them (“We’ve decided to give you 1 year of vest, OH, and years 2, 3, and 4 as well!”) and Scott saying, “A founder goes down with the ship.” The applause from our people seemed to last forever, and I was on the verge of tears myself.
Now comes the real work. We are officially one company, operating as two entities, but with much to do both internally and externally. We have work to do on both our team and on our product, to realize our shared ambitions. It was a wild ride this week, and now the story has been told. We are excited to move on to the next chapter. We’re putting our heads back down to focus on building the future of AI together.
Ryan Castner retweeted
many researchers i know have become jaded and mostly rely on word of mouth to find papers
i've been doing the usual @latentspacepod conf coverage and still hold a candle for the low key, serendipitous finds
this year it was @sumeetrm's paper reporting that
1) ai's can collude/evade human knowledge via steganography with schelling points (low-no coordination, statistically undetectable steganography via @casdewitt)
2) it is a @_jasonwei emergent property, improving with scale
3) he has also worked out multicast/decentralized collusion as well
we recorded the whole paper presentaiton for the podcast recap but heres the paper
Ryan Castner retweeted
2min after i said this my guy ran this
aws rds delete-db-instance --skip-final-snapshot ...
maybe don't let my guy into prod yet
Ryan Castner retweeted
today we are announcing reinforcement finetuning, which makes it really easy to create expert models in specific domains with very little training data.
livestream going now: openai.com/12-days/
alpha program starting now, launching publicly in q1
Ryan Castner retweeted
The product trio is merging. And it's not just because of AI.
In fact, I first saw this concept in 2020, presented by Yuhki Yamashita, CPO of Figma.
(ie, well before the current AI boom)
Ryan Castner retweeted
every time I go to a swe’s house, I steal their copy of DDIA to see how long it takes them to notice it’s missing. I’ve done this over 20 times and nobody’s caught on. That’s why the modern web is trash! Nobody reads this book, they just buy it and post it for clout.
Ryan Castner retweeted
Please share this far and wide. As far and wide as you can. NIST Password Guidelines for 2024 are in the process of being updated.
This is a HUGE pet-peeve of mine (when vendors in particular are still operating like its 2017 and keep changing passwords every 60 days, STOP DOING THIS, it's outdated and has been shown to put you MORE at risk than less -- NIST explains why it does in this document, meticulously outlining user behavior**) so I'm sharing this in the hopes all of you will pass it along to your bosses.
The Special Publication series governing passwords is SP 800-63 "Digital Identity Guidelines".
The 2024 version is 800-63-4.
Here: pages.nist.gov/800-63-4/
The companion docs are also on that link. They are 800-63A, 800-63B and 800-63C. These are different documents for different scenarios in play at your org.
The previous update was in2020.
The changes in the 2020 version from the 2017 version were numerous but one of them was that the password verification method should NO LONGER require passwords be changed at specific intervals (i.e. every 60 days) but in the following circumstances instead:
1. After a breach/compromise
2. User request
2024 repeats this and adds a bunch more guidlines but here is a screenshot of page 13 of the new 800-63-4 (note the # 4 after it) which outlines how your systems should now and moving forward, be handling passwords.
This goes for Active Directory, too. All your systems which have passwords should align with these guidelines provided there isn't another standard or framework you must adhere to which overrules this.
Most frameworks, however, have moved away from arbitrary password resets and complexity rules.
**We cybersec researchers and hackers use wordlists from breaches in a variety of different ways. Hackers use them in tooling to crack passwords whereas researchers use breach dumps to see the kinds of passwords users are creating and the psychology behind them.
Using complexity rules gets you the user psychology of:
Password1
Password2
and so on
Use phrasing instead and allow for spaces, which is important. Humans type phrases with spaces. They also mention phish-resistant methods and most vendors are on-board with MS going to be turning off all Legacy Auth next month, across all free accounts and tenancies.
I'm so excited for the new changes!
Ok I'm off my soapbox.
Share the love! Thank you!































