motoboi

> He’s also in the conversation for a Nobel.

How does that work, actually?

I believe the whole string vs stringbuffer that later was made redundant by compiler contributed to that vision.

People started dismissing allocation discipline as a thing from the past because "that thing was solved a lot ago and the compiler now is smart enough".

Well, for string, yes, but not for arbitrary objects.

Oh, god. They are marketing it as an old people artificial friend. Probably will blame users in the future when they get attached and have ai-induced psychosis.

Disgraceful.

On the technological side, it's a marvel!

Nothing much smarter than run the agent in the project folder, make sure it has means to interact with production (like configured kubectl) and ask it to jcmd the hell out of it.

Rust? OK.

C++ or go? Then you'll have to take a very closer look, because the java JIT is wonderful. A masterpiece of several hands, actually.

I found claude and GPT very helpful on this, because java have a very sofisticated monitoring harness. Just ask the agent to connect to the running application (on kubernetes or whatever) on prod and do a java flight recording then analyze allocations.

I managed to improve some applications of ours from several garbage collections per second to several minutes between collections. That _really_ improves p99.

Gemini, gpt and fable are actually very good compressions of internet content. But is lossless compression as in they kept the most important part (for them to fulfill the next token task) and found a way to mimic the rest.

bean in mind that "GPT‑5.6 Sol on Cerebras at up to 750 tokens per second" not necessarily means the same model (in terms of inference result). It can mean anything like a very quantized model, a different level of model activation per inference etc.

Of course we can trust that wouldn't name the same thing with different levels of intelligence, right? Right?

This is actually a very nice question and the answer is that interpreted languages with a JIT benefit from this.

One example is Java, which will happily vectorize your code into AVX or SSE where possible.

Python just got a JIT compiler and we’ll start seeing the same thing soon.

But as someone else said here, some constructs don’t translate well and adding transformations to show vectorization would negate the perfomance gains.

Sad that the compiler (even Java) can’t explain you this and warn about it, but now with LLM, maybe they’ll start doing things like that soon.

A json schemaless stream querying engine that would run several sql queries over the same kafka consumer (not a consumer for each query).

Yeah. We grown being trained to solve those small puzzles that are websites and apps, so we learned _how they are projected_, not how they work.

I mean, we learn that a enroll is normally a flow. Flows have steps. So if you came to the end of the flow and the finish button is gray, you think.

Hum… I'm used to flows. This is a multistep flow. Flows normally need me to fulfill some small checks and won't let me proceed between steps if something is missing. But some won't. Maybe this is one of those? Some flows have warnings in the end, some have next to the thing missing. I don't see any warn in the last screen, so I'll go back every step and check field by field for errors. That'll probably do.

This is the model you have in your mind, of how a website or an app works.

People that came to computers, apps and websites later in life didn't learned the puzzles.

Yeah. And reflect on the fact that you know that a 404 error is not a form error. Old people won't. Heck then won't understand that a chrome error page is not your site.

Some people know how to fix a fridge.

Yes, but even non-technical users have a different relationship that formed over the young (neural plasticity) days.

"I click the <next> button on this form, nothing happens. Given that I have an internalized notion of forms as a multistep flow, maybe it couldn't advance because something is missing on the current step? Maybe it will give me those messages below or above the form? Maybe some read message somewhere? I'll search for those and try again"

vs

"I click the <next> button on this form, nothing happens. Not sure what would happen next. Maybe the thing will tell me? I can't see no dialogs, no messages, no error screen. Nothing changed. Maybe if I press it again? Nothing? Hum, maybe is disconnected from the network? It's the wifi again? Maybe if I power cycle the wifi router it will reconnect? I click next again now I get a clear error message: no internet connection. I suppose my internet is broken again. Will call my tech-savvy friend."

Old people. They exist.

Not even that old. 60 year people can't user your fancy site because then don't have an internal model of how a computer works.

You know that when pressing a button a hidden engine runs in the backend (or something runs in the backend). You expect an answer and if the expectation do not match the result, the model in your mind creates an hypothesis about what maybe happened and iterate from there. Maybe you should have clicked something before? Maybe you should mark some form checkbox?

Old people don't have that because they didn't grow up with computers.

What is on the screen is what they see. I clicked next and nothing happens. Well... the site is broken.

You known when you plug your refrigerator and nothing happens and instead of reflecting on the possible blown out resistor that you can bypass with a small wire you understand that your only relationship with the refrigerator is plug and unplug or call for help? That is an old person using your site. They won't fight against it. They'll give up immediately.

Not sure, but in my experience, instead of asking for code, i'm asking for solutions and providing a kubectl configured to reach my cluster and az monitor command to read the logs and telemetry.

A typical session is the agent establishing a metrics and log baseline, creating the code, compiling, deploying, observing, fixing, redeploying, observing metrics, determining the outcome and commiting.

I really, really, don't look at the code anymore.

UPDATE:

so my point is: it won't have my stewarding the code anymore, but it will have the infrastructure (and ultimately the real world) providing feedback on the traces.

There’s nothing much new about the architecture. The real gains come from the usage traces.

It turns out that having a text based interface for a text-trained model creates a very nice feedback loop.

Right now as we speak, people are generating text traces on anthropic and OpenAI servers that teach their models to do everything under the sun, text wise.

So people right now getting super mad at how dumb the model is when reverse-engineering a super complex function from binary, when they write “stop, you dumb robot, you are going wrong, go this way thank you very much” are actually leaving a lesson in the form of the "chat" text history.

Some may say that each bad word get us closer to ASI.

That and obviously the order of magnitude more efficient GPUS we got that allow for different tradeoffs at training time.

Very nice. But nowhere to be seen in my model list on github copilot enterprise ai settings? I suppose it's still rolling out. The "rolling out to github copilot" is verbatim on the blog post, not my words.

On the other hand, opus 4.8 became immediately available at copilot and foundry when launched.

Mai-voice-2 and mai-transcribe are now available for me on foundry though. Just half a day after launching.

Hear me out: i love microsoft. It's sad to see this state of AI business.

Unless you are token rich, you'll have to find a way pretty soon.

For tasks (like kubernetes, linux, reports, database exploration and such) I use GLM5.1. Faster is actually smarter in those cases. And much cheaper too.

Opus 4.8 is for the unknown. Things I don't know how to do myself.

To understand microsoft IA problems right now, observe that NONE of the models announced are available for use even in the microsoft foundry, which is the place were you add models to your account.

I understand github copilot rollout takes time, but why can't we consume the models via microsoft own api after launching?

Anthropic models are available at foundry the same moment they are launched, but not Microsoft's own models.

For me the greatest difference is that superset is terminal-centric while conductor is chat-centric.

GUIs slow you down, in my opinion. But having the nice visual diff is something we can't really do well in TUIs, so very welcome.

So, superset, for me (been using for quite some time now) is basically to organize my agent and terminal sessions per task and project.

I can switch context much easier and can also resume working on something days later, with all my tabs nicely available and separated.

This was consuming me before, a dozen or more tabs and windows in my computer that I don't really remember to which task each belongs to.

I strongly believe you don’t need to call another model for that. The same model can do result fine. Just not as part of the same context.

I mean that if you ask codex on gpt 5.5 to submit to a plan reviewer subagent that uses gpt5.5, this is enough to have a very good reviewing and reassessment of the plan.

My hypothesis is that it’s even better than opus.

The reason why submitting the product of one LLM to another to review is that you need a fresh trajectory. The previous context might have “guided” the planer into some bias. Removing the context is enough to break free from that trajectory and start fresh.

The microbiota is passed from mother to son on birth, not totally from the environment.

What we currently don’t understand is why for some people they never got them (we have techniques to transport the biota from the mother during birth for non-natural procedures) or they loose them.

Even with the transplant, the microbes won’t stick around on those people (not taking about autistic people here, but people in general).

Diverse food really helps, just as not eating ultraprocessed (they won’t reach the end of the intestines).

Fermented and other pre or probiotics will really help too.

But none of those will recover the biota in some people.

Not apples for apples.

Before:

- Opus 4.6 each premium request is 3 premium requests

After:

- Opus 4.6 each dollar spent is 27 dollars in copilot AI Credits.

Given that you'll receive 19 dollars of AI Credits in Business plan, that means you can probably say 1 "hi" to opus per month.

Wait until you discover how many wrong labeled images in imagenet and that it still kickstarted the deeplearning revolution.

It’s saying that 16% of the problems have well, problems.

Please next time start with azure foundry lol thanks!

This is how anthropic captured the code agent so fast. You need training data, users are giving it to you.

Being a terminal application, all interaction is trainable signal (unlike, say, cursor, which is an IDE and let users freely explore, edit the files, move the mouse. Model sees nothing of it, nothing to train upon).

So meta is doing the obvious, we want to train a computer use model, we need training data. Better to capture from employee than buying low quality data.

Well, the model that a lot of people have been given access to and are reporting about on twitter?

With the previous generation? Yes. With 10T mythos-level models? Not even close.