Activity for engineer.sitereliability.best

Me daring my kid to write in cursive, 2026.

The Verge@theverge.com

For comparison www.theverge.com/tech/979583/...

A lot of the harnesses today are jumping to concurrency for speedups because LLM responses are relatively slow. As faster inference providers come online, I wonder if we see single-threaded harnesses on fast inference outperform the swarmy systems on some tasks.

“Of all the marvellous achievements of modern science, the Electric Telegraph is transcendently the greatest and most serviceable to mankind. It is a perpetual miracle, which no familiarity can render commonplace.” Hold my beer.

apenwarr@apenwarr.ca

Ok this is very neat

o no, the janky toy harness got complex enough i prob want a real workflow engine. this way lies madness.

One of the important things an eng interview should try to figure out is "does this person understand that outcomes at scale get very weird and unintuitive?"

dax@thdxr.com

deepseek is insanely good at inference - they are hitting 96.56% cache ratio on our heavy traffic second best provider for us is doing 91.60% this doesn't seem like a lot but it means they're using ~2x less GPU time

Love digging into "why did this agent go off the rails?" and finding the answer was a subtle word choice in a single prompt 4 conversations ago. A large part of working w/ coding agents is developing a spidey sense for when you're about to monkey's paw yourself into a big mess.

LLMs will have good theory of mind for writing apps when they understand that wiring up a connection heartbeat api to the client doesn't mean visibly displaying heartbeat pulses to the user.

No GPT, it's not cool that you circumvented 'task: deny' on the harness by calling bash -c "workmux" to spin up 4 claude code instances to do the work for you.

Off Protocol Live — AMA with Bluesky’s Head of Platform Jim Calabro

In the past I said that software consultancies will be the first to make hay of the capabilities of AI models and my friends at SWE consultancies are telling me "no, we're getting our butts kicked." Instead of updating on that my response for now is, "yeah you actually have to deliver 5x more work."

OK who's building time tracking on atproto?

Emelia@thisismissem.social

Good gosh, Harvest is having a laugh after being bought by Bending Spoons. €116,50 for a monthly plan, that's billed monthly, not yearly. That's a 807.8% increase. There's gonna be a hell of a lot of unhappy customers getting that email. I think I've worked out how to keep my costs manageable.

I don't think I'm being cynical here and throwing up my hands and saying "nobody has good security so this is normal and ok." It's that our normal way of building system where security is a top priority is to make it ok to ship very slowly. I don't think that is the right call for the frontier labs

Alex Wise@engineer.sitereliability.best

I think I suffer from this more than many but "big company w/ incentives to move quickly on tech innovation claimed its security practices were better than they were" does not update me in any way. It's most of why A\'s position on safety has always rang hollow. The incentives run the wrong way.

We're going to be looking back at the LM architectures of this era like this.

Sung Kim@sungkim.bsky.social

They find that training a single transformer layer can recover most of the gains achieved by full-parameter RL training, and in some cases even surpass it, instead of a systematic layer-wise study of RL training. Paper: arxiv.org/abs/2607.01232

the feeling when you click like before you read the second sentence of the post.

Maybe we should improve the open IP pathways so if Discovery Loop starts casting off a bunch of useful but difficult-to-monetize discoveries they can be easily made available to the general public.

Something like half of bluesky posts are not in English. There is no "on bluesky". It's just recsys algos. Ask attie.ai to make you a feed if you aren't seeing what you want. Use labelers to mute accounts you don't want to see. It's your experience to own.

Matt Hodges@matthodges.bsky.social

lol

In 2007 Adobe updated the branding of Illustrator to be "Ai" instead of "AI". Would love to follow the thread of the universe where that didn't happen and see what that mess looks like for their design team.

Anyone who has adminned an Exchange server for a moderately-sized company knows execs will be the worst customers for automated email handling.

Alex Wise@engineer.sitereliability.best

Effective assistants are one of those things that can be useful now but will be hell to productize for the mass market. Whether I can reliably sort your email depends a lot on the distribution of your inbox and how much you care about otherwise innocuous emails.

I know everybody loves the idea of native speech-to-speech models, but SST->LLM->TTS pipelines let you be *very* clever in the turn loop and now at pretty respectable latencies.

Sung Kim@sungkim.bsky.social

Google DeepMind's DiffusionGemma Technical Report They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of