tanvach

So you're saying, you don't mind having cancer?

The biggest shift, in my mind, is that there will no lower bar for morality. Companies goals do not need to be aligned with society to trade for labour. The only lever we have is to block new data centers from being built.

This is like saying cancer is perfectly normal.

That’s really cool. I think you nailed it, English isn’t the best way to encode my intentions in code. A pseudo language is likely much more efficient.

For me, LLMs tend to engage the 'language center' that drains me faster than the 'problem solving center' I usually reserve for writing code. We really need a different abstraction the bridges the gap between human and programming language, and load balance between these two parts of the brain more effectively.

This is very apt

Seems like Localsend doesn't currently work reliably with Tailscale enabled. This is a bummer. Hope they also allow sending files between clients on the same tailnet, that'll be super neat.

Super cool, lock-in is very real. We are overflowing with Duplo and Lego sets because I just don't want to deal with another system. There are, of course, other models on Thingiverse, Printables, etc., but knowing these are properly designed to fit and work is a huge plus. Cudos to the team!

I think it’s not right to blame teachers, because you cannot control that right? Agree it’s far too hard to find aspirational teachers, but they are naturally in demand and will be snatched up by well paying positions, like private schools. The tool is supposed to help given the current suboptimal teaching practices.

It’s human nature really, and we see this in our jobs, instead of using LLMs to learn to do our jobs better, we replace work with automation.

It’s going to be quite hard to motivate students to learn now that they know answering can be automated.

Was about the mention this, -CC has been working perfectly for me

Super interesting, thanks for sharing

Anyone has done mesh WiFi (ideally triband) using off the shelf parts and Linux?

I have an Orbi AX system which works reliably, but now I want to upgrade the radio to WiFi 7 and that means I need to upgrade all the hardware.

Hoping to move to using off the shelf parts so in the future I can just change the radio (ideally bunch of USB sticks).

I understand this is not strictly just the router. I can (and used to have) a router as separate device, but any mesh WiFi right now that I can find need a pricy router that acts as the coordinator, essentially negates the economic benefits.

That distinction is obviously quite apparently from the context. I’m using atomic clock to be consistent with what these radio clocks are marketed as.

That’s a very good point but the transmission power is extremely weak and won’t propagate more than 1m.

Related - we have an atomic Seiko wall clock expecting to have the time automatically adjusted by the WWVB LF atomic clock broadcast. Turns out, the signal is very weak where we now live. Manually setting the time on these atomic clocks is a HUGE pain (beware!).

Turns out it's possible to emulate the atomic clock signal quite easily with a Raspberry Pi, or in my case I put together Arduino code that can emulate atomic clock broadcasts from around the world using an ESP32 module using NTP servers: https://github.com/tanvach/clocksync

The history of these atomic clock broadcast signals and their differences in different countries is quite fascinating.

Long computing embedded in confusing texts should be sufficient.

It might work for crude actions, but my take is that touch signal is nuanced and won't get transmitted via mechanical linkage well. Love to see it prototyped though.

Before, I was humbled, learned and mastered.

Now I feel frustrated and numb.

I’m not an AI skeptic, but more and more I feel the joy has been sucked out of doing anything that AI has touched.

Indeed :) Ah yes I agree! That was awful to everyone (and to analyze!).

During the pandemic, we had a natural experiment at my previous company. Our org had started an org wise auto ‘meeting starts 5 minute past’ while others had the traditional meetings start on the hour.

Also conveniently, we also had the calendar data for internal meetings, internal VC software (not zoom) db that logs the participants when they join and leave meetings and employee function db.

I was serendipitously the lead DS for analyzing the effectiveness of the ‘starting 5 minutes past’. After joining and cleaning a lot of the data, the data showed:

1) at the start of the trial, meetings ended on time. Then after few weeks it slip to ending late, negating the usefulness. Other orgs did not see meetings running late. 2) ICs tend to stick around and over run meetings, while managers tend to leave meetings on time. 3) if I remember right, we had a survey data that showed pretty clearly that managers prefer the ‘starting 5 minutes past’ while ICs do not care or have negative sentiment.

The biggest predictor for people who prefer starting late is how crowded their schedules are. Managers tend to have very crowded schedules which means they want a break between meetings, while ICs prefer not having to waste time waiting.

In the end we reverted back to normal schedule. It was just easier for busy people to bounce early.

People looking into this a little too much, looks to me like random walk. You should try reinitiating the trial (or have multiple running) and see if the ranking is robust.

Coincidentally we were on the Robotaxi during the black out (didn’t know about it, we were going to Japan town from the Mission). Noticed that it navigated through the non-working traffic lights fine, treated it like a stop sign junction. One advantage of building unsupervised system from public version that had to deal with these edge cases all around the country.

Though the safety driver disengaged twice to let emergency vehicles pass safely.

A plug for tidyverse adjacent data.table - really should be combined someday :)

A little meaningless with one-shot, should try recreating a few times per model and see what the variance looks like.

If you really care, I would suggest helping with documenting how the process should work for others to reference going forward.

How so? Can you describe more of your experience if you don't mind sharing?

I find it more straight forward to just model the failure rate with the variables directly, and look metrics like AUC for out of sample data.

Agreed, these type of analyses benefit from grouping by cohort years. Standard practice in analytics.

The abstract reads legitimately bonkers