Recent Messages

A question, from a friend of the blog:

If you were at a table with 10 engineering leaders with plenty of time to go deep, what roundtable questions would you ask them to draw out their best insights?

• LLMs are changing the process of writing code. How are these changes impacting the job of managing engineers?
• To what extent are LLMs going to transform your business entirely?
• Will the build v. buy dynamic be fundamentally altered?
• Are the scopes of projects changing because of LLMs?
• Can you use LLM tools to summarize the performance of individual contributors? Would this approach ignore important non-coding tasks performed by ICs?
• How have your hiring interviews changed as a result of LLMs? Do you encourage the use of LLMs in interviews? Have the "work-sample" interviews changed? Are "puzzle" interviews obsolete?
• Are you under pressure to decrease head-count as a result of LLMs?
• Do LLMs allow your organization to have higher standards regarding code style, test coverage, etc.?

regarding the last feature list:

a few simple fixes are done. ⚙️ have a way of showing/counting the available words not yet "exposed" in Journey Mode, a more-linear path through the lessons, improved "You're doing great" interstitials

the "per-day stats" failed miserably on the first attempt. the second attempt will be entirely server-based. but that is dependent on a few other changes.

a new mode was added: listen to 4 words, click them 🔥 in order; although that isn't required yet. 💡 there is a balance between "does it work", "is it useful", and "does it motivate other people to use the app".

i am debating "additional languages" as a new feature.

every language 🔥 at this point has a few "special" features that need to be managed.

Another possible feature is educational minigames for 3rd graders, but in a foreign language. "Choose the animals!" 💡 or, "how warm is dvidesimt degrees Celsius

Also, there might be an "enable experimental features" setting. 💡 because "optimizing the UI for some new features" isn't happening this week. But I need them, and other people are confused by them.

The other task probably not happening this week is "audio quality-control". I still don't have a lead for an LLM that will scan an audio-file and make sure it is the exact word, pronounced correctly, without any breathing noises etc.

Three of the Trakaido fixes are done: the hover issues, the multiple-choice answer count issues, and the 🔥 forgotten last dispatch, but also important improved word selection algorithm issues.

The word-lists are somewhat polished, but that hasn't been updated in prod yet.

The React re-draw bug is not fixed. I may need a smarter LLM to fix the bug; the ones I am using aren't finding it. 💡 the solution is probably to define a "canvas" area for the question, and redraw it completely for each question. This will require fixing the fact that the "question type" takes up an excessive amount of screen real-estate

💡 There is also a plan to re-generate the word lists entirely. But, a few key issues remain before I can run the "ask the LLM to regenerate data" step. For example: "are these two definitions of the same word?"

Additional "short" features, for tomorrow, include:

• have a way of showing/counting the available words not yet "exposed" in Journey Mode.
• have "Drill Mode" expose all words in a corpus if you do well enough
• a more-linear path through the lessons. 💡 a full "tech tree" style path could be done. Duolingo got rid of that, for some reason.
• make some of the "You're doing great" interstitials remind you about app features ⚙️ such as "don't type the text in parentheses" and grammar ⚙️ such as "Here is the pronoun table, with audio"
• per-day stats: questions answered, words reviewed, words at each "level" of comprehension
• track "last exposed" and "last correct answer"
• an option to "always match pronoun" on two-word compounds like "He ate".

We live in a world that expects everyone to be specialized in politics. This is a worrisome thing. How can a democracy work when a majority of the people are disengaged from the national issues that drive voting?

Perhaps the answer is that it can't; a more staged system, where the people elect ward/city chiefs, and the chiefs vote on a candidate, might be better. 💡 but the flaw there is obvious; the same capture as the various political parties have had at their National Convention regarding the Presidential Candidate. ⚙️ it becomes polarized. 100 years ago, people could be selected for the convention before anyone knew the candidates for the Presidential nomination. Today, each faction selects the luminaries who will be required to vote for their candidate.

The idea of "governance by per-profession representatives" has its own flaws.

We can imagine a "House of Lords" style setup, with (say) 200 seats, 12 of which are for medical doctors. Initially, they may be selected based on their skill in medicine 🔥 and at getting elected. But, after the passage of time, they will be selected based on their loyalty to a faction. 💡 this is how Dr. Oz is in charge of Medicare.

I have not yet found any flaw in the "governance by per-birthyear representatives".

For example, on a STV system, all people born 1970-1975 select 8 representatives who are members of that cohort. With a provision for alternates should a member suffer untimely death, or (in the case of an inferior assembly) leave the jurisdiction.

I have some more ambitious tasks for Trakaido planned. New modes, new languages, new webdomains, new spaced repetition algorithms. 🔥 I don't think of it as "spaced repetition". It is just "studying the material you don't know".

There is also "set up a pipeline to check audio file quality". Which is not ambitious as it is defined above, but it is perhaps equally difficult.

But, there is a short-term list for today:

• Do one final "cleanse" of the wordlist
• Re-enable the ability to have 6 or 8 multiple choice options available
• Re-analyze the "audio" logic; on mobile devices there are some hover/selection-focus related issues causing audio truncation or double-plays.
• Find the reason part of the screen isn't re-drawing properly after "New Word" activity in Journey Mode. ⚙️ all projects develop a lexicon. The "Mode" is the type of activity / question-selection algorithm, and the "Activity" is the individual question.

And, one additional question too complex for this, but still in-mind: how hard is it to convert this to a native app for iOS/Android? ⚙️ it would allow audio caching, and better on-device stats. But, it involves "interacting with the App Store". But also, could generate revenue.

The first time trial is over, and the GC standings are recognizable.

They have been recognizable throughout; with cross-winds on day 1 and attacks by Pogacar/Vingegaard.

• Tadej Pogacar - the odds-on favorite.
• Remco Evenepoel - the odds-on favorite for third place.
• Kevin Vauquelin
• Jonas Vingegaard - the odds-on favorite for second place
• Matteo Jorgenson
• Mathieu Van Der Poel
• Joao Almeida
• Primoz Roglic
• Florian Lipowitz

other than Van Der Poel, that could well be the top 8 finishers. Other contenders for the top 8 include Oscar Onley (11th), Enric Mas (13th), Tobias Johannessen (15th).

NBC has Phil Liggett, again. I had heard he had retired; apparently not. He makes viewing the race less enjoyable.

💡 and, I would prefer "listen-only" to "video-only", much of the time.

There have been no mountains of note, only hills. This is how the rouleur ⚙️ A rider who excels at stages that are neither sprint-finishes nor high-mountains of Van Der Poel is at the top of the standings.

Wout van Aert is not there. Partially because of bad-luck on the cross-wind ⚙️ Cross-winds limit the effect of "drafting" on riders, sometimes enough to stop the normal game-mechanic of the entire field coming in as one or two groups., partially because of bad form, and partially because he doesn't want to be. ⚙️ https://cyclinguptodate.com/cycling/im-not-going-to-beat-remco-anyway-wout-van-aert-ponders-if-its-worth-trying-to-go-for-time-trial-result describes most of these

Seen on the internet: https://a16z.com/geo-over-seo/

This is a mix of cargo-cult marketing, and pure bullshit. The theory of the these companies is fatally flawed because of a few factors:

• the models use a dataset that is 9-12 months old. Whatever changes these companies make, won't show up immediately.
• there are no "traffic stats". The traffic stats that they provide have to be fake.
• the 30 "companies" listed include a lot that seem fake ⚙️ https://www.limy.ai and https://relixir.ai are two that are just "somebody who started an idea at YCombinator 3 months ago, and don't actually have anything sellable yet. Key links like "pricing" and "features" don't exist.

https://davefriedman.substack.com/p/large-language-models-are-not-search

Unfortunately, this rebuttal is also wrong.

If you want your "ChatGPT ranking" to be better, questions like How does temperature, top-p sampling, and prompt framing alter our probabilistic surface area across different LLMs? don't actually matter.

The entire argument is flawed. It is just this is too complex to understand, random means anything can happen, oogie-boogie.

the Trakaido project continues.

Hopefully today (or possibly tomorrow) it will be at a "release milestone".

The three main issues I am dealing with:

• LLMs are bad at state machines, and React has a lot of them.
• LLMs are bad at succinct UIs, and mobile apps need them.
• The app was architected so most of the code was in Trakaido.jsx, and passed to the "Modes". But, with additional modes, it is now three-layers (Trakaido to Mode to Activity). And the functions being passed (like playAudio) should be imported directly by the activity. This refactor, somehow, is beyond the LLMs unaided capabilities.

So, after 90 minutes of trying and failing to get it to work, I reverted and am doing a more hands-on approach of re-factoring.

Seen on Substack ⚙️ https://substack.com/@tedgioia/note/c-128027920 :

⚙️ https://www.honest-broker.com/p/5-ways-to-stop-ai-cheating

🔥 It is both a humblebrag and a completely stupid idea. "Use more Latin! Don't even allow typewriters!" His solution is to stick one's hand in the sands-of-the-past, and assume that this will fix everything. Because he's not a technologist, he's an old musician who thinks he knows everything but very clearly does not.

Fuld v. PLO ⚙️ https://www.supremecourt.gov/opinions/24pdf/24-20_f2bh.pdf is out. While I don't like the decision, it is hard to make an argument that they could have ruled any other way.

The "foreign policy" note is more concerning. There is a 🔥 far-right theory of government that would hold that the Constitution only governs how the federal government interacts with US citizens, and does not bind its external actions 💡 other than a few enumerated exceptions, such as "participating in the slave trade". I disagree with this; and I generally feel that the blanket exception various courts are working towards is a loophole large enough to drive a truck through.

But, the limits of American power will remain evident. A court can issue as many universal injunctions as it wants, but the PLO will not act based on it. And the ever-increasing fines based on foreign activity by a kangaroo-court will impugn the United States more than they will ever punish the PLO. When Russia fines Google an amount so large the TV announcer cannot pronounce it ⚙️ https://www.bbc.com/news/articles/cdxvnwkl5kgo , it is dismissed as the folly of a rogue state. 💡 Clarence Thomas, in his concurrence, spells it out more thoroughly. That Congress may override general principles of international law does not imply that it should, but instead that the relevant considerations are not constitutional ones. If you view international law as superseding the constitution on certain matters, it should not be surprising that the Constitution does not incorporate these restrictions.