HN user

garspin

77 karma
Posts5
Comments29
View on HN

1) Making money & keeping money are 2 different skillsets. You've made some $$$, now learn how to keep it.

2) Time is far more valuable than money. If you can take life-changing $$$ off the table in exchange for time, do so. The 2nd $1M buys you a tiny proportion of the benefits that the first $1M did.

3) You have a v. high risk concentrated portfolio that is aligned with your income. That's massive risk.

4) Taking it now buys you time & optionality. Leaving some still buys you blue sky. Best of both worlds.

I doubt it. Brains run at only a few operations per second.... GPUS at TFLOPS. There just isn't enough bandwidth.

My brain only needs to get mugged in a dark alley by a guy in a hoodie once to learn something.

GPT-4.5 1 year ago

and then you found that it offered marginal-but-not-amazing improvement over the previous version.

Then call it GPT 4.1 and allow version space for the next iteration.

I think the label V4.5 is giving the impression of more than marginal improvements.

An alternative explaination...

Deepseek is a side project for a hedge fund.

Shorting NVIDIA & releasing everything including the source would have a high probability of being hugely profitable, with almost zero downside if it went unnoticed.

Joseph Heller (of Catch-22 fame) was one of them.

His response to a billionaire was 'I have something you will never have - enough'.

LLMs are text compression algos. They are very good at memorising & retrieving text related to the user input. They can even pass bar exams based on that - some misinterpret that as intelligence.

However, no amount of scaling will change a text memorisation algo into a symbolic reasoning or composability algo, both of which are necessary for progress towards AGI.

So yes, LLMs are a dead end in the quest for AGI. However, they have their uses as a google/stackoverflow replacement.

Flying for months at a time in the stratosphere, Zephyr offers game-changing capabilities that will be transformative for mobile connectivity and earth observation.

As a payload agnostic platform, Zephyr can transform into a multi-functional tower in the sky to provide low latency 5G direct-to-device mobile connectivity services.

Hardest to diagnose phone support I had to deal with was an elderly relative in the DOS days trying to find another lost file....

"type in 'dir \documents' and press Enter. What comes up?".

"Nothing".

"Nothing at all?"

"Nothing happens".

After trying many variations of this, it eventually became apparent that he was pressing the Space Bar instead of the Enter key.

Surely consciousness is not black & white, but a scale. Rocks at one end & us(?) at the other.

There was a recent post suggesting that our thoughts are 98% unconscious.

It's not hard to imagine that there could be some animals with more than 0.01% consciousness. After all, we started at 0% conscious and evolved a little - other species are probably some way down that track.

mind-bending D3

That's something that bugs me too. D3 is pretty amazing but is only accessible to coders.

we’ve almost completely lost the art of offscreen rendering

markdown2.com is something I'm working on to address that. It lets you design markdown documents containing charts & diagrams, and then Save As HTML/PDF or SVG/JPG. There's also an API for generating them in bulk.

The point Marcus is making is that LLMs are solely next word predictors - intelligence requires far more than that.

LLMs will always have the same basic challenges - their 'kinda copy what a human has written' algorithm has hit its limit.

He is an advocate of symbolic learning - getting that to work is a lot harder, however it's far more likely to result in genuine intelligence.

For many products/services I use a disposable email/login (from duck.com) - the site gets some data, but it's wrong (DoB 1 Jan 2000), anonymous and unrelated to any of my other disposable logins.

Obviously doesn't work for anything that needs real phone# or real physical address, but anything to reduce risk of data breach, cross-referencing data, tracking helps.

I also use disposable digital credit card number (from wise.com) for non-mainstream sites.

It's a lot easier than having to navigate the 'delete all my data' maze.

I too tried DL & learnt too slowly & gave up on the irritating gamification.

I run alone (at a slow conversational pace) for 5-6 hours a week, and see that time as an opportunity to learn conversational German. Any ideas what I could be listening to... maybe a listen & repeat type app ?

as they simply truncate the body of the article, leaving you no choice but to open the original url to read the rest of it

Inoreader has a 'Load Full Content' button above every (shortened) feed item. It retrieves the original in full & displays it in clear text (no ads and little html formatting).

I discovered the button by accidently clicking it - it's an unintuitive steaming coffee cup.

I've switched to self hosted wiki.js for my blog & personal notes. And pocketbase as a headless CMS. I self host it - however free hosting is at pockethost.io