How do you learn something outside your domain? My favorite way to learn complex concepts is to write them in my own words. The book Writing to Learn shows how to do this for any field. Anyone?

With the help from LLM, when can we shorten the travel time to Mars from 9 months to 9 weeks or 9 days? People seem to worry more about losing jobs, but there are so many exciting problems to work on.

Wild story: the Python semantics Anthropic used was defined at Brown by @JoePolitz and students from a *MOOC*. Summary here, which also explains what Anthropic depends on: blog.brownplt.org/2013/06/18… Funnily, the last sentence says, "Once again, machines may put people out of work".

As a $TSLA shareholder, I think it makes sense to merge the two companies.

It feels so good to solve problems with Ops 5.5. It is not just smart, but also like a real person on the other side!

The first thing Opus does is to enhance the error variant in the enum. Now we're talking!

and one more slop from the cursor model: My code has defined state for the iCloud sync, final class SyncActivityMonitor { enum State { case idling, uploading, downloading, uploadingAndDownloading, error, disabled } Now to show the detailed error information, what will a good engineer do? Turn the `error` enum to carry error information: error(ErrorInfo). What does the LLM do? It creates another struct struct SyncStatus: Equatable { var state: SyncActivityMonitor.State var failureReason: String? } Now we have two sources of truth to maintain! You can't be this bad!!

This AI slop is so frustrating, Cursor grok 4.6 xhigh. I asked LLM to display the iCloud sync error of the NSLocalizedFailureReason in a popup. I gave it the errors in the first picture. It tried twice, the first time it tried if let reason = nsError.localizedFailureReason, !reason.isEmpty { } A decent attempt. After I tell it the `nsError.localizedFailureReason` field is empty, guess what it comes up with? a BFS, walking down all keys in the nsError.userInfo field. It even detected back edges to prevent infinite loops for this BFS (the second picture). Don't ask me how I feel when I see this code. What do I expect it to do? It should notice the error was propagated from the NSPersistentCloudKitContainer.eventChangedNotification, which contains only the container event -- it did not contain the actual reason of sync failure; it had the error "The operation couldn’t be completed. (Cocoa error 134400.)" -- the container sync operation couldn't be completed, and that was because of the auth failure. The 134400 is a good indicator of the auth problems at the higher level, from there, it should add another observer to the auth state! That's not the end of the story. Then to solve this problem, the LLM treated the 13440 something special, and wrote code to enumerate all types of status for the code 13440. Yes, of course, I reverted it (see pic 3).

So a 27b model is half of the price of grok 4.6. Either grok is highly subsidized or the 27b model is too good. If the latter is true, consumer GPU with 12GB/24GB vram and Mac/MacBook 24GB/32GB will sell off soon. Interesting time.

LMAO

Mark has been cooking

Jeff is leaving Google!?

When it works, sunshine and moonlight streaming works perfectly fine, but when it doesn’t, it is really painful to use: mouse missing, iOS pointer and remote mouse both appear but out of sync, remote server discovery failures. Need something better.

I realized I have not posted anything about changing my window manager on my Linux desktop yet. Changing a window manager is usually not a big deal, but managing windows is like managing the multiplexed terminals, and I do see a significant productivity increase with Niri (I have switched from Omarchy to CachyOS) Here are some thoughts: The factorial spiral splitting behavior in almost all tiling window managers is flawed in design when adding new windows. It significantly changes the original window size for both horizontal and vertical splits when there are more than two windows on the screen. It is bad because it makes the text jump. I often find myself going back and scrolling it back to the place. The scrollable WM lets me append windows on the right without changing the existing layout; they stay where they are. This is superior because all old window and text layouts are retained. But scrolling alone is not enough when windows accumulate, since random jumps become harder. To address this issue, one more must-have is the overview. It lets the user zoom out and see all window tiles on all screens. This solves the quick-jump problem; I can immediately find out where my Slack is when the layout grows too long with the overview, and the Slack window does not change whatsoever.

Some random writing today: In 2018, a bit more than one year after I joined Meta, we have expanded the POSIX storage service to about 10 regions with only 3 engineers in the team. The on-call rotated to me every 3 weeks. I loved it because I saw so many production issues in the distributed file system (split brain, slow self-heal, high tail latency, sporadic latency spikes etc.). For that half a year, I grew so much because of the on-call. Then my team doubled from 4 to 8 people in a few weeks. All I remember now is that in the first few weeks after the expansion, I mostly wrote onboarding tasks and did code reviews all day because the new hires were really good. They completed onboarding quickly, and the team review policy required a first comment on a diff within a day. The tasks were carefully chosen and written to give new members a full picture of our work. I split ongoing projects into stages, wrote tasks for each, and included recommended implementations plus potential improvements. Retrospectively, it was such a unique experience: not just writing tasks, but a strong way to build a team. At lunch we discussed the tasks and their history, and chatted about the future of the services and the team, so our new friends learned a little about everything every day and we learned about each other too. More importantly it also showed strong care for new members. I respect them. I want to help them. The tasks aimed to grow them. Almost all of them, including all my future interns, showed great appreciation for the details in those onboarding tasks when they first received them. I wrote those tasks hoping they would grow like me to care for and improve the service together. Again, retrospectively, when I just joined the then-Facebook, the production engineering team mailed a Kindle and books on the Linux kernel, computer networks, Unix system programming, and more. I was touched because somebody cared about my growth. It was all about caring and growing together. Receiving that care made the team feel like friends, not just 9-to-5 coworkers.

Opus 5 is up. RIP Opus 4.8, now gone.

Meanwhile, I am working on minimizing tokens, as much as possible.

I have been thinking about the saying that “LLM is like a compiler” for a while. I like everything about compilers and know a thing or two about their internals. With knowledge of operating systems, I have a coherent view of the whole system when an ELF binary is loaded by the dynamic linker. I’d like that coherent view of LLMs too. That sounds exciting!

Fable: You’re right — the only reason … Me: Uh-oh

Composer 2.5 is such a great model for execution at its price range.