prpl

4.8 at 10k tokens a second would be more groundbreaking than a Opus 6 IMO.

Terence Tao did NOT say that. Amit Sahai did. Maybe you are just being less careful?

I’d happily take Opus 4.8 xhigh at a reliable 10k tokens per second over oversubscribed SOTA models.

At some point this will probably happen, but we’re not quite there yet.

a few things are global (login service) to some extent, just as AWS places several such things in us-east-1

We’re talking about ~’94 here…

Not just gamma rays, cosmic rays too, although they filter the UV and are targeting nitrogen fluorescence and catch the cerenkov overlap. I’m sure there’s many people here who have worked on them.

I worked on HiRes/Telescope Array in Utah (Not CTA).

Not sure the thrust of your comment, but what is really going on is just novelty seeking to get a bit of a dopamine rush, which slides into creativity.

It all comes down to dopamine at the end of the day.

A SOTA model at 10k will be materially different, even the model is 6 months old.

I have been working on a cross platform way of sandboxing without a VM:

https://github.com/brianv0/formwork

sure it’s not tricky. But everybody does it different and OpenAI couldn’t even be bothered to do it right when benchmarking their models

For one, I’ve been working on a generic sandbox environment

github.com/brianv0/formwork

You should be easily able to hide/lock down files, network, and MCP tools from an agent and it shouldn’t be up to the agent.

It’s been a very long time, but I read Alas, Babylon in middle school.

I remember aspects of their modern existence also being problematic (IIRC a house in florida with lots of windows doesn’t fare well without AC)

I’ve been formalizing behaviors to requirements and building more end to end tests to verify them. In reality, every requirement needs a verification element. Secondly, building traceability in your unit/integration tests helps “keep it honest” about how much it has built and verified.

Mostly I’m borrowing heavily from model based systems engineering and adapting it to some form of constitutional spec driven development.

My impression is that many of the indian engineering schools are some form of trade school and it shows. A lot of people are very good at churning out things, but they vastly prefer to use well defined tools for their trade and do a lot of work impassionately. Taking the creative path or even reflecting on failures isn’t natural.

On the other side of things, you have PhDs coming out of the US with too many ideas and the inability to deal with basics such as sprint planning and CI/CD.

Not sure if this will help, but I would love it for the Colleges of Sciences peoples instead of a postdoc.

Yeah you could go with just Number though, and then keep almost everything a BigInteger/BigDecimal under the covers (or dynamically choose the correct class)

Are you being serious or sarcastic?

I built a 40 (later 80) node cluster with clusterknoppix ~2006 to run a bunch of physics simulations off of old library computers after I replaced a bunch of PSU fans. Kept my cubicle toasty until we moved them into a random subterranean room I think was used for some early nuclear research at university.

Salesforce runs Salesforce on postgres with LSM:

https://vldb.org/cidrdb/2026/a-multi-tenant-relational-oltp-...

you've had an account for four years, and _this_ is your comment?

SAT isn’t an IQ test, and probably all sorts of people took it before they had cultural awareness or a diagnosis that would have lead to different testing conditions had it been taken after diagnosis, let alone the fact that test scores are not comparable.

GPAs similarly not comparable over large time ranges, schools, or degrees without normalization you can’t get.

there’s really no limit to how many times and ways you can review something with AI, except dollars.

They are not all filtered early nor are they necessarily missing from images. The data can be embargoed until orbits change effectively.

Right, people haven’t internalized that these are really just scripts in natural language.

Oracle is the hilarious version of this -

9i - "internet"

10g - "grid"

11g - "also grid"

12c - "cloud"

26ai - "ai"

various other examples. One really annoying thing is this has also happened in open source projects too - generic things that, sure, help out with AI tasks are now "AI" things.

"rely" is overly strong in these cases usually (more like "make use of")

> The real core of infrastructure monitoring isn’t dashboards. It’s the alerts.

“it’s not X it’s Y”

at this point when I see this pattern in writing I assume most if not all of it is AI generated - same with em-dashes.

This is not to discount the idea that alerts are more important than dashboards (I work directly in observability) - but just to say that I personally shut off reading anything else with these patterns because, generally speaking, the rest of the content is just not original or interesting.

Can you build most of this with Ray?

For files up to 100kB of size, this should effectively be really close to the same price as S3 when writing (didn't check reading so much, but the writes/PUT is always much more expensive than read/GET)

Would be really useful pre-compaction and to deal with small files issue without latency penalties

Adversarial software development is also when I do my best work

I have honestly not had that problem much. Being specific, concise, and strong with your prompts helps out a lot.