supersour

Thank you so much for this! You have helped me considerably.

Maybe a Show HN? I would be quite interested in seeing the results of this project

100%. I am really getting tired of the narrative that code before LLMs was optimally performant, perfectly architected, completely understood, and bug free...

Does the author not have the experience of working in a legacy codebase that nobody really "understood"? Something sufficiently complex where even the senior SW devs needed to scope out project work and research the codebase for dependencies or potential issues?

I fail to remember a time at LARGE_CORP where even the most experienced developers were able to scope out or design a feature without studying the existing documentation, timing diagrams, etc....

100% agree. I understand the argument around the "fun" of programming and learning. But my joy from coding (especially when working 9-5, on a team of 5-500 people) always came from the value creation for other people and for society.

AI has simply been a tool to accelerate that and I've genuinely never been happier about the value I provide to society. I find it hard to take the "respect of the craft" argument at face value - even if I enjoy knitting by hand, I can still support the existence of mechanized looms and machinery that provide affordable clothing in the millions.

My late grandmother knit 1-2 pairs of socks per day and donated them to the homeless center - absolutely amazing thing to, and it kept her motivated and sharp up to 104 years of age.

However, if her ultimate goal is providing socks to the homeless, would it not be better for her to work a library job for $15/hour, then take the $100+ a day and purchase then donate 30 pairs of socks? Even if the socks are a lower quality and wear out twice as fast, she could help 15x the people thanks to the mechanization and efficiency provided by the devision of labour.

I see this divide more and more and truthfully I wish I could understand both sides. I am encountering trouble in the startup I'm working on for exactly this reason: I feel laser focused on the attempt to create value for others, while my mech eng cofounder has this "respect of the craft". From from my perspective this has only caused delays in generating revenue.

Any help in understanding those who "love to make" (as opposed to those who "love to deliver") would be extremely useful.

Yes, power efficiency may be the largest benefit of these chips actually - especially if the projections are true that the US and other countries simply aren't able to ramp up power generation to meet forecasted datacenter demand.

There's a bit of a ticking time bomb there, something that a Taalas-like architecture can clearly resolve.

Yes this is true to some extent. I've been using LLMs to run some computer vision tests, and I've certainly noticed myself running into the trap of "just one more AI experiment", or "just one more change" while neglecting to actually properly integrate the learnings into my mental model.

However, without an agent running its own experiments on a cloud GPU, would I realistically have invested my limited work hours and tried evaluating 10 different models, each with 10 different tuned parameters, to solve my specific use case?

Or would I have tried 1-2 models and spent my time trying to optimize those models?

I think there is some merit to the spray and pray approach when one is in the exploration phase of the solution space.

Also, on more than one occasion now, I have had fable halve the inference latency of a model simply because the original implementation from an academic included unnecessary GPU-to-CPU-to-GPU transfers or similarly inefficient operations. Those optimizations came at essentially 0 time cost to me and I can verify that the outputs are byte-identical. Pretty sweet!

A property developer is the other real-world example which comes to my mind.

Yes this is true, but we actually have a lot more data to back this on than exclusively fMRI analysis - for example the ADHD medication guanfacine works only because the alpha 2 receptors happen to be wired differently in the prefrontal cortex than it is in other areas of the brain (a2 is inhibitory for most brain regions, but in the PFC they're positioned to amplify connections between neurons) , so by stimulating alpha 2 we allow for a more “top down” control from the prefrontal cortex than we do without, which improves executive function.

So that is one extremely robust way to understand neurological conditions like ADHD or Parkinson’s

Hi, do you have a way to get in touch so I can ask you a few questions about your lab? Thanks

If this article resonates with you in even the smallest way, I urge you to read Neil Postman's "Amusing Ourselves to Death: Public Discourse in the Age of Show Business".

I am currently re-reading this book and am amazed by the apparent accuracy of his analysis, which is that the mediums in which we communicate or express information (print vs. TV vs. TikTok) have a massively understated role in the quality and type of communication we participate in. That is, as print lends itself naturally to logical argument and less to emotional knee-jerk reactions, the type of conversations taken place in long-form print will by nature be more logical and intellectual. Compare this to TV or short term videos, which captivate us using more primal forms of distraction (bright lights with moving images, fast talking, "Gotcha" type rhetoric, cool dances, sexual/romantic behaviour, or background subway surfers), and it is obvious that the nature of what we see is inherently less based around logic and reason.

And as a consequence, if we are what we consume, it is only natural to surmise that the quality of the mind follows the quality (and qualia) of our media.

I think the prior probability in the bayesian sense is that the two entering cohorts are equally skilled (assuming students were randomly split into two sections as opposed to different sections being composed of different student bodies). If this were the case, the implication is that performance differences in standardized tests between cohorts are due to the professor (maybe one of the profs didn't cover the right material), so then normalization could be justified.

However if that prior is untrue for any reason whatsoever, the normalization would penalize higher performing cohorts (if it were a math course, maybe an engineering student dominated section vs an arts dominated cohort).

So I guess.. it depends

I'm not familiar with GPU architecture, is there not a shared L2/L3 data cache from which this data would be shared?