Hasnain says:

I still need to get back to reading the postmortem on the recent sycophantic changes to ChatGPT. But damn..

“In that sense, Westgate explains, the bot dialogues are not unlike talk therapy, “which we know to be quite effective at helping people reframe their stories.” Critically, though, AI, “unlike a therapist, does not have the person’s best interests in mind, or a moral grounding or compass in what a ‘good story’ looks like,” she says. “A good therapist would not encourage a client to make sense of difficulties in their life by encouraging them to believe they have supernatural powers. Instead, they try to steer clients away from unhealthy narratives, and toward healthier ones. ChatGPT has no such constraints or concerns.”

Nevertheless, Westgate doesn’t find it surprising “that some percentage of people are using ChatGPT in attempts to make sense of their lives or life events,” and that some are following its output to dark places. “Explanations are powerful, even if they’re wrong,” she concludes. “

Hasnain says:

Looking forward to future posts in the series.

“For organizations dealing with similar challenges, FixrLeak offers key takeaways:

Prioritize structured code analysis: AST-based techniques help ensure fixes are safe and context-aware.

Automate targeted fixes: Focus on well-scoped, high-confidence fixes first to maximize success rates.

Integrate AI responsibly: Validate AI-generated code with rigorous testing and code review processes.

While FixrLeak is currently deployed at Uber, the principles behind it—combining static analysis with GenAI—can be adapted to other large-scale systems. Companies seeking to automate code quality improvements can explore similar techniques using AI-assisted code repair, AST analysis, and structured prompt engineering.

This blog kicks off a series from the Programming Systems group on leveraging GenAI for software engineering challenges. Stay tuned for the next post, where we explore how GenAI can automatically fix data races in Golang”

Hasnain says:

Bookmarking for later reading.

“This paper presents a tool called RustAssistant that leverages the emergent capabilities of Large Language Models (LLMs) to automatically suggest fixes for Rust compilation errors. RustAssistant uses a careful combination of prompting techniques as well as iteration between an LLM and the Rust compiler to deliver high accuracy of fixes. RustAssistant is able to achieve an impressive peak accuracy of roughly 74% on real-world compilation errors in popular open-source Rust repositories. We also contribute a dataset of Rust compilation errors to enable further research.”

Hasnain says:

This is going to give me nightmares as I sleep because what the heck, man

“So to put a bow on this:

The o3 model isn’t smoke and mirrors, tricking us by only using EXIF data. It’s at a comparable Geoguessr skill level to Master I or better players now (at least according to my own ~20 or so rounds of testing).

Humans still hold a big edge in decision time—most of my guesses were < 2 min, o3 often took > 4 min.”

Spoofing EXIF data doesn’t throw off the model.

Whether you view this as dystopian or as a technological marvel - or both - you can’t claim it’s a parlor trick.”

Hasnain says:

Great work by some great folks here, gonna bookmark this for late re-reading

“In some instances, the LLM resorted to “cheating” by producing patches that superficially resolved the issue without addressing the underlying problem. This can occur when the generator modifies or removes code in a way that prevents the crash from occurring, but does not actually fix the root cause of the issue. We observed that cheating happens more frequently when we request the LLM to retry within the same trajectory. A potential solution to this could be to empower the LLM to say “I cannot fix it,” which may come with a tradeoff with success rate. However, note that most of the cheating was caught in the verification step, highlighting the utility of differential testing.”

Hasnain says:

If I was still at Meta this would probably go in the clowntown group

"Personally, I use a solid color background. It was the default in Windows 95,¹ and I’ve stuck with that bluish-green background color ever since. It’s sort of like my comfort food.

Imagine my surprise when someone pointed me to a support article titled “The Welcome screen may be displayed for 30 seconds during the logon process after you set a solid color as the desktop background in Windows 7 or in Windows Server 2008 R2.” Why is logon slower with a solid background?"

Hasnain says:

We must bear witness. I did not think I’d be agreeing with “the American Conservative” but here I am.

“As the days roll on and the death toll piles up, how will future generations remember the role we played in this disaster? Will they view us, and Trump, as peacemakers and not the ones that looked away? It is not our responsibility or in our interest to save the world, true, but it’s also naive to believe that we do not play a sizable role in permitting Israel’s continued bombardment and starvation of the Palestinian people. Something must change and quickly or the blood of innocents will forever be stapled to the recorded rule of Trump and the MAGA right.”

Hasnain says:

““God did the State of Israel a favor that Biden was the president during this period, because it could have been much worse. We fought [in Gaza] for over a year and the administration never came to us and said, ‘ceasefire now.’ It never did. And that’s not to be taken for granted,” the former Israeli ambassador said.”

Hasnain says:

Gotta love fun bugs and UB

“This was the most interesting bug I’ve encountered for a while. I initially had a hard time believing that a bug like this would directly tie to a specific OS release, but I was proven completely wrong. At the end of the day, it was a simple bug in San Andreas and this function should have never worked right, and yet, at least on PC it hid itself for two decades.

This is an interesting lesson in compatibility: even changes to the stack layout of the internal implementations can have compatibility implications if an application is bugged and unintentionally relies on a specific behavior. This is also not the first time I encountered issues like this: regular visitors might remember Bully: Scholarship Edition which famously broke on Windows 10, for very similar reasons. Just like in this case, Bully should have never worked properly to begin with, but instead, it got away with making incorrect assumptions for years, before changes in Windows 10 finally made it run out of luck.

Yet again, we are reminded to:

Validate your input data – San Andreas was notoriously bad at this, and ultimately this was the main reason why an incomplete config line remained unnoticed.
Not ignore the compilation warnings – this code most likely threw a warning in the original code that was either ignored or disabled!
In the end, the GTA players are lucky: in many other games, issues like this would’ve remained unfixed and they’d become a folk legend. Thankfully, GTAs are moddable and well understood, so we can act upon problems like this and ensure the game stays functional for many more years to come.”

Hasnain says:

Ooooof. I have heard people talk about these group chats and elite cabals before, but oof. Especially that screenshot at the end.

“The tone was jesting, but “Marc radicalized over time,” Hanania recalled. Hanania said he found himself increasingly alienated from the group and the shift toward partisan pro-Trump politics, and he came to see the chat he’d established as a “vehicle for groupthink.” (A friend of Andreessen’s said it was Hanania, not Andreessen, who had shifted his politics.) The group continues without him.

Hanania argued with the other members “about whether it’s a good idea to buy into Trump’s election denial stuff. I’d say, ‘That’s not true and that actually matters.’ I got the sense these guys didn’t want to hear it,” he said. “There’s an idea that you don’t criticize, because what really matters is defeating the left.” He left the group in June of 2023.”