2y of data. They should address that confounder and demonstrate the results are stable even if you stratify over time. 2026 models are very different from 2024 ones, after all.
I really liked the main question. But I’m more interested in the structure how we journal, not if you use Logseq, OrgMode, paperback, etc.
For example, I try to journal by treating my work as “experiments” and tie those in turn to goals and challenges. And my experiments consist of a plan, prediction, the actual data, and the evaluation (of the data against the prediction). Finally, I constantly track the current state relevant towards the goal. If that sounds familiar to the Toyota Kata, that is because it is. And it should sound familiar to anyone trying to apply the scientific method.
For Mixture of Expert models (like GPTs are), they can produce different results for an input sequence if that sequence is retried together with a different set of sequences in its inference batch, because of the model (“expert”) routing depends on the batch, not the single sequence: https://152334h.github.io/blog/non-determinism-in-gpt-4/
And in general, binary floating point arithmetic cannot guarantee associativity - i.e. `(a + b) + c` might not be the same as `a + (b + c)`. That in turn can lead to the model picking another token in rare cases (and it’s auto-regressive consequences, that the entire remainder of the generated sequence might differ): https://www.ingonyama.com/blog/solving-reproducibility-chall...
Edit: Of course, my answer assumes you are asking about the case when the model lets you set its token generation temperature (stochasticity) to exactly zero. With default parameter settings, all LLMs I know of randomly pick among the best tokens.
To me it implicitly means: “I am selling you my distilled copy of the book at nearly the price of the book. That lets me recoup my investment, and after a few days I will take it offline, so I don’t get a law suit from O’Reilley for copyright infringement.”
Oh, wow, scummy to scammy, indeed! It’s all in the fine print… I didn’t bother reading this appendix on the first read of the blog post.
Disclaimer: I do really enjoy this book, as it reminds my of uncle Bob’s Clean Code, but a short version, and with the focus on what to do after writing code, when you need to change it or want to understand it better.
Yes, it is structured very much like a collection of blog posts. Which is great for me, as I typically work on learning one “tidying” a day or less. So I am not yet done with the book, end-to-end.
That’s an interesting viewpoint, even surprising to me. Can you be more concrete about what you think is wrong with the agile software development movement Kent Beck co-founded, and how it destroys developers?
I agree on principal, but believe it is more nuanced and hybrid can work if you implement it uniformly across the company. For example, I’ve seen places where everyone can WFH Tue & Thu or Mon, Wed, Fri where hybrid have a quite smooth collaboration culture. But if it is “random”, every day ends up being WFH - that’s in effect just the same as fully remote, and you need to act accordingly. And if you work in a global company with globally distributed teams, the company already is de facto fully remote, and the picture becomes even more nuanced.
I use this. You can easily pick dates (and it correctly handles time changes in one region vs the other) and you see the time bands side by side. Finally, it has a great mobile app, too.
Here’s another example: Apple Arcade decided to discontinue my kids’ favorite local multiplayer game (Marble Knights). Apple erased the game from my devices and there is no way to get this game at all now, at least as far as I know. With the pending increase in Arcade’s monthly costs, that was the last straw for me to say goodbye to this Apple “offering”.
Just to expand on the above, as two days later I now feel it is hard to understand how my reply is related to imtringued‘s last paragraph (“if you can fix the asymmetry or eliminate the zero lower bound it doesn't actually matter what form your money takes”):
If you don’t allow banks to “print debt” out of thin air by eliminating fractional reserve banking, as Daly suggested in his work, I believe you eliminate that asymmetry. Actual money is lent, actual money has to be returned.
The question that remains then is if it is valid to ask for interest on that money. But I believe that is acceptable, and the rate might be set to 0% or even negative in the right circumstances.
Thanks for the pointer to the Leitner system (my last name coincides…) But why would you want to use that simplified Leitner system if Anki can track each card individually?
Disclaimer: I don’t know if Anki supports a simplified system that applies spacing to a whole group of cards instead of each card individually, but why you would want that?
I don’t think Daly is suggesting to move away from debt as an instrument in his work, only to establish a 100% reserve system for commercial banks, including the Fed. He was suggesting to nationalize money supply via the Treasury, to avoid bank runs and fiascos like they happened just now to CS and SVB.
Apparently, this was already the direction of thought among some economists in the 1920s (like Frank Knight and Irving Fisher), but then got overruled by the eternal growth mindset that evolved during the Great Depression (which Daly claims leads to “uneconomic” growth). Printing money as an instrument should very much be available - but only to the Government/Treasury.
I am currently reading the late Daly’s 2014 book “From Uneconomic Growth to a Steady-State Economy” that seems to address exactly your concern for a more fairly distributed economic system. A real eye-opener that would require us to change three things to build a more sustainable economy and future:
(1) stop fractional reserve banking and nationalize money again to avoid the very quantitative easing/tightening fiasco we are seeing today,
(2) measure economic progress with a quality of life index instead of GDP to accurately reflect the cost of products/resources and put more emphasis on life quality improving services, and
(3) replace the current international trade rules and system run by the IMF with a more equitable system where capital costs are accurately reflected at the source to avoid “outsourcing” problems to third world continues.
While all three changes are difficult to achieve in our current society, I am amazed at how clearly Daly’s life work can outline how we could build a much more sustainable economy and healthier society.
Maybe his amazing last summary book is of interest to others on HN.
Rejoice, as you have issues that are really easy to improve upon.
First of all, don’t ever argue about solutions with your peers (e.g. whitespaces), and dive deep (Can you imagine that it is hard for others to read your code? Have you tried reading your own code that you have not touched for 5-10 years? Etc.). From what I hear, you might just ask the question to yourself before you confront your coworkers. And then ask Why at least several more times for every answer.
You also are very close to understanding the core issues: It typically matters to understand the business problem and find the simplest solution, with the least effort to resolve any problems.
You might get good mileage out of studying domain driven design at this point in your path.
I prefer System76, because it ensures that hardware and software work together flawlessly. Most threads here seem to complain about that, but I never had an issue on that front with my Oryx Pro.
Only hardware complaints are that it can become noisy with (Nvidia!) GPU use and that the battery life is short (at least if you are used to MBPs…) But for example, I have had no serious headache with the Nvidia GPU, only some driver update juggling.