sudhirb

Is this not the good kind of problem to have? Subnautica 2 is doing so well that the developers get their earn-out bonus?

Seems like pure profit-maximizing/greed!

I can hardly believe my eyes! I helped do some related research specifically concerning thin-film drainage from tubes, way back in my undergraduate days: https://doi.org/10.1016/j.expthermflusci.2018.04.015

a 90% saving is huge isn't it?

for long agent sessions, I would expect a very high cache hit rate unless you're editing the system prompt, tools, or history between turns, or some turns take longer than the cache timeout

So long as perceived LLM skill is still "spiky" - e.g. within a domain, still showing relatively high variation in ability (often depending on the task or user, to be fair), people will continue to dismiss it

In the general case I think this is a great idea - if we do a good job of documenting intent etc. in commit messages, agents have an easier time understanding why lines of code exist with no additional specs/mechanisms/etc.

Interested to see what techniques in this area pull ahead and gain traction!

Coding agents are such a congested space right now that to me this mostly reads as an advertisement.

I have 150Mb/s FTTP for £37/month - upgrading to gigabit would be £75/month, for example!

Interesting that non-salty water didn't make the string conductive(?) enough - I'd have thought that there might have been enough soluble stuff in string.

Also I believe this person runs the ISP I use (and I couldn't speak more highly of it - Andrews & Arnold).

For me, a lot of the draw is that it's cheaper than managed db services for small/toy projects of mine (that I don't want to use dynamo db for) - that and in a previous job it was useful as relatively temporary multi-tenant storage.

The partner for these projects has a benchmark that the top frontier LLM labs seem to be running on their new model releases - I think there's _some_ value to these numbers in helping people compare and contrast model performance.

https://andonlabs.com/evals/vending-bench

I think that "manipulate people for financial and political gain" is an outcome of what social media companies actually do - I was under the belief that in a general sense, they want to maximise the time people spend on their apps so that they can sell this attention to advertisers, independent of whether or not a given ad buyer wants to manipulate people.

Some mild whataboutery: is the purpose of a cancer ward to fail to cure a large fraction of its patients[0]?

https://www.astralcodexten.com/p/come-on-obviously-the-purpo...

Both miracles are illness-recovery related and feel to me quite like regression to the mean, but I can imagine this is somewhat of a strategic move from the Catholic church to bring some relatability into things.

For me, the USP Warp used to have was generating shell commands from prompts inside the terminal - but Cursor has had this in its embedded terminal for a while now so increasingly I find myself using Ghostty instead

the icann wiki has some articles for these:

https://icannwiki.org/.agakhan

https://icannwiki.org/.ismaili

https://icannwiki.org/.imamat

> One of the consequences of this is that we should always consider asking the LLM the same question more than once, perhaps with some variation in the wording. Then we can compare answers, indeed perhaps ask the LLM to compare answers for us. The difference in the answers can be as useful as the answers themselves.

There was once a coding agent which achieved SOTA performance on SWE Bench Verified by "just" running the agent 5 times on each instance, scoring each attempt and picking the attempt with the highest score: https://aide.dev/blog/sota-bitter-lesson

Anecdotally I think I have heard "what all" most commonly spoken by Indian English speakers - though that's probably quite far outside the scope of this site.

>Ah yes. An unverifiable claim followed by "just google them yourself".

Some agent scaffolding performs better on benchmarks than others given the same underlying base model - see SWE Bench and Terminal Bench for examples.

Some may find certain background agents better than others simply because of UX. Some background agents have features that others don't - like memory systems, MCP, 3rd party integrations, etc.

I maintain it is easy to search for examples of background coding agents that are not Jules or Copilot. For me, searching "background coding agents" on google or duckduckgo returns some of the other examples that I mentioned.

It appears that e2b runs Firecracker microVMs (https://e2b.dev/blog/how-manus-uses-e2b-to-provide-agents-wi...)

It shouldn't be too hard to get a Firecracker orchestrator running locally - the articles here were very helpful when I was doing this myself: https://jvns.ca/blog/2021/01/23/firecracker--start-a-vm-in-l...

By no means are better background agents "mythical" as you claim. I didn't bother to mention them as it is easy enough to search for asynchronous/background agents yourself.

Devin is perhaps the one that is most fully featured and I believe has been around the longest. Other examples that seem to be getting some attention recently are Warp, Cursor's own background agent implementation, Charlie Labs, Codegen, Tembo, and OpenAI's Codex.

I do not work for any of the aforementioned companies.

I've worked somewhere where CORBA was used very heavily and to great effect - though I suspect the reason for our successful usage was that one of the senior software engineers worked on CORBA directly.

I have a biased opinion since I work for a background agent startup currently - but there are more (and better!) out there than Jules and Copilot that might address some of the author's issues.

I am suspicious that the buffalo mozzarella registers as "tangy" at all - though I suppose it travelled quite a long way

I remember being somewhat sold on this story by the PinePhone, but it seems like it might not be possible to buy one new nowadays.

Having just looked up the PinePhone again for the first time in a while, it does look like the Ubuntu Touch project is still alive and kicking, and compatible with some modern commercially available phones!

The main thing preventing me from having a non-standard Android phone/distribution as a daily driver is access to mobile banking apps - I'm yet to check for myself but as I understand, having an unlocked bootloader means that banking apps will consider the device "compromised" and will not work.

I've had an ancient Parker 51 for a good decade or so now which I use almost daily, that originally belonged to my great-grandmother.

I'd expect there to be a reasonable amount of variation in how long these pens last due to differences in usage, machining tolerances, ink types and materials - though mine has done very well considering how many times I've chucked it into a bag, dropped it on hard floors, etc. (I've probably just been lucky so far).

I think the budget play is to get a cheap refillable fountain pen and some cheap fountain pen ink (I bought some Diamine bottled ink about 10 years ago and I've still got plenty left).

More expensive fountain pens are indeed luxury products - replicas are often available on places like aliexpress.

I think it probably feels like bullshit if you don't think you'll gain any satisfaction or derive any pleasure from the act of writing or the aesthetics of your writing instrument.

For coding agents, evaluations are tricky - thorough evaluation tasks tend to be slow and/or expensive and/or display a high degree of variance over N attempts. You could run a whole benchmark like SWE Bench or Terminal Bench against a coding agent on every change but it quickly becomes infeasible.

I'm not sure when I would rather rebuke someone for sending me AI-generated output than either just ignoring it, or sending a polite minimal response to the same effect

I think:

SFT = Supervised Fine Tuning TTFT = Time To First Token TESCREAL = https://en.wikipedia.org/wiki/TESCREAL (bit of a long definition)

"on ine itabalism" = online tribalism?

Interesting! I also wonder if jj helps coding agents to be better able to resolve merge conflicts which I've anecdotally found can be painful with the current crop