psyklic

These have been on HN:

- https://www.bookai.chat/

- https://www.myreader.ai/

Even if LLMs are thought of as Markov chains, there are still substantial unsolved scientific questions.

For a given prompt/"state", an LLM essentially computes the next-state probabilities. This is done by compressing language and storing particular patterns/distributions. To date, we still don't understand what statistics are stored by LLMs. Or even what stats might be necessary to produce natural-sounding language. (A canonical Markov chain only works with n-gram statistics, but we know these are insufficient.)

IMO figuring this out to a human-level understanding could be a major breakthrough in science. It would reveal a deeper structure to language. It may even help understand how our brains process language; we'll at least know one plausible algorithm. Prior to LLMs, many scholars assumed language was unique to the brain and could only poorly be "computed".

Last week I trained an LLM to auto-unrotate a Caeser cipher, without it knowing the rotation key. It turns out there are only a few possible ways this can be done, e.g. by analyzing letter frequencies. Of course, the LLM must use one of these algorithms! Determining this high-level algorithm would be "understanding" what it is "actually doing".

Through analysis, I proved it indeed counts each letter's frequency. It then combines the "overall" letter frequencies with another distribution of "first letter" frequencies. This increased its accuracy to nearly 100%.

To me, this description is much nicer than had I "followed the code" and described it as millions of multiplications and additions. (Much of this ends up being noise anyway.)

It sounds to me like you really care about mentees being successful and are investing a lot of time/effort. I believe you can change your outlook a bit into knowing it's moreso on them to flourish. For example, you sound extremely hands-on for a mentor. I would not do all of their coding for them.

IMO it's not your job to teach how to use a framework. But instead of assuming they "should" know things, you could assume they are skilled but at different aspects of development. So from this perspective, a mentor might give additional topics to brush up on to be successful in the job. And still be super open to providing internal career advice, helping with company-specific things, etc.

It sounds like maybe a lot of what you write here is important for mentee success yet not being communicated. You could probably be more upfront about team expectations, and let them know what a typical dev workload will look like.

Hiring is tough, and some people are not able to perform as well as expected. That can be frustrating, but it's not your fault. You may want to clarify how the company evaluates new hires so you know the process.

As a student, I say just try everything you can and decide for yourself. Mathematica is basically a huge collection of already-implemented scientific algorithms. So, no more hunting for Python packages that may or may not work. The main downside is most people outside academia avoid proprietary languages -- and it's expensive for commercial use.

I'd say out of the box Mathematica has ... nicer visualizations, more available functions, better symbolic math, etc. It might be possible to get some similar functionality in Python, but it would be a hassle. On the other hand, Python will have the bleeding-edge latest implementations since more people code in it.

That said, Mathematica does keep up to date. For example, the latest version auto-imports neural net/LLM weights, visualizes their structure, and lets you capture values from anywhere in the circuit. IMO it's great for learning how they work. And it has some neat ChatGPT integrations, auto-runs code, etc.

- This is my favorite intro Mathematica book (for programmers/academic-minded people): https://www.amazon.com/Mathematica%C2%AE-Problem-Centered-Ap....

- There are also good free tutorials like their Fast Intro for Programmers: https://www.wolfram.com/language/fast-introduction-for-progr...

Agreed -- slightly getting the name wrong does a national/international search! Searching for coffee shops often leaves out obvious results such as some Starbucks. Even more frustratingly it doesn't show particular businesses even when I zoom in, forcing me to manually search for the business name.

Also: It would be wise to verify their professional references. Diligent VCs will check them when you start pitching, and your co-founder's poor references will hinder you from raising capital.

In grad school, I (briefly) researched synthetic DNA logic gates. Using these, you can build neural networks in DNA [1] and probabilistic switching circuits (my co-authored paper) [2].

To do this, you design DNA strands that bind to each other in designated regions when mixed.

To make a logic gate, you just make some bindings conditional on other bindings. To detect that the gate is closed, you make the final binding uncover a fluorophore, which is detectable by a machine. (This is all called a *displacement cascade*.)

Using our techniques, the DNA bindings could not be undone. However, I imagine that a source of new DNA strands to replenish the old would effectively implement re-callable functions.

[1] Neural nets in DNA: http://qianlab.caltech.edu/nature10262.pdf

[2] Probabilistic switching circuits: https://www.pnas.org/doi/full/10.1073/pnas.1715926115

Empirical studies are only useful until the system is mathematically understood. For example, I can construct transformer circuits where the skip connection (provably) purely adds noise.

I can also prove in particular cases the MLP's sole purpose is to remove the noise added from the skip connection.

You could perhaps hire an improv coach for a team workshop. Level one improv is all about building team rapport and helping individuals feel more comfortable contributing in a supportive group.

I'm of the opposite view - publish only if you can prove it. If you reward half results, the field will be deluged with half truths. If someone publishes a hypothesis you're working on with only a half proof (and stands to gain the future credit), then there is little incentive to continue doing the work to prove it.

This was actually a major issue one of my PhD advisors had, since it led to poor foundations for the field with little incentive to ensure their validity.

Was at Handmade Seattle a few weeks ago and saw a (still early beta) Android editor called MobileCode - https://play.google.com/store/apps/details?id=dev.mark.mobil...

Lets you download GitHub repos to work offline and has common coding actions in a quick tap menu.

You might enjoy the book "Exercises in Programming Style (Second Edition)" by Cristina Videira Lopes. It looks at how you can compute term frequency using 40 different coding patterns, explaining the how/why of each.

I organized a tour here for a RustFest conference in Barcelona. Submitted the form on their website, and luckily someone was willing to show us around. Highly recommended!

Choose the one you're most excited about learning, then go all in as deep/energetically as possible to stand out. The determining factor isn't the job market; it's yourself becoming demotivated when the going gets tough.

Being neutral could also mean striving to make decisions based on independent standards, as opposed to a bias toward one side. For example, a judge who bases decisions on evidence and the law is typically considered neutral -- even when they rule guilty.

My conclusion was that if you're young/healthy, HMO vs PPO may not matter as much. That said, I went with a PPO because in my case: (1) it's faster to see a specialist, since a referral is not required; and (2) in my area, it has vastly more in-network doctors who are highly-rated.

One gotcha I've run into is that most plans (both HMO & PPO) are only valid in a single US state. The one exception is "emergencies," but when I asked it seems this is basically only the ER (and I was told only if it turns out it's actually an emergency, whatever that means).

I believe you can get out-of-state covered via out-of-network. But as mentioned, the out-of-pocket max on my plan is shockingly ~6x more for out-of-network than in-network. So you won't pay 90% of $200k, but it will still end up a very substantial amount. But maybe that could still be worth it to see an out-of-state specialist?

I think Steve is approaching this from much more of a business angle. Looking at the Business Model Template slide in the article, programming language choice may not affect the business model much, except perhaps in rare cases where the language/tech achieves a business moat. Founders can definitely hold random theories about the business market/strategy, as long as they are willing to invalidate them (profound beliefs that are loosely held).

Google Maps is similar. Google clearly stores your Home and Work locations, labels them on the map, and allows you to change them in the settings.

Yet with "Web & App Activity" off, when asking for directions it states: "To see your home and work ... turn on Web & App Activity."

u know it ;)

Could also consider an M1. Here's a new M1 MAX 64GB RAM & 4TB hdd for $2999 -- https://www.bhphotovideo.com/c/product/1668351-REG/apple_mbp...

In nearly everything, you can do some "minimum amount" that only makes the appearance of getting by. This works until it doesn't (e.g. the minimum threshold is misjudged). Eventually, others see through it and opportunities come less often.

The best students do not constantly compare themselves to slackers. Instead, they compare themselves to those at the top of their game. So there is awareness their GPA could potentially end up minorly lower, but taking easy classes simply isn't an option.

> The lawsuit goes on to say the $100 is the price the co-founder originally paid for the shares when the company was formed.

My former business partner tried to pull a similar stunt. He mailed me a check for a few dollars (the original purchase price), along with a cheery matter-of-fact explanation. He assumed I'd cash the check and in doing so give away my vested shares for almost nothing, but luckily I did not. Sent a letter back explicitly stating I continued to own the stock and that the shares were not eligible for repurchase under our repurchase agreement.

If you pursue this, make sure to research the tax implications of giving away large chunks of equity post-founding.

If you wait until the company has a large valuation, your co-founder could be taxed on the stock's current value, even if they aren't making enough salary to pay the taxes. This is why many founders purchase the equity up-front (at a very low par value) then do reverse vesting (where the company can claw back the equity under certain scenarios).

A social network's utility scales exponentially with size (https://en.wikipedia.org/wiki/Reed%27s_law). The front end can decay, but people will still use it since everyone else is on it.

This is especially true with dating sites -- why even try a small startup's dating site when you'll get way more matches on Tinder?

What makes it even more difficult to leave incumbents is that the industry has basically found a few replicable ways to effectively hijack attention/dollars. IMC has integrated them into each of their dating apps (Tinder, OKC, Match, etc). For example, one of the first hit dating sites was Hot or Not, of which Tinder is essentially a mobile phone version.

This post suggests that Awair might not retain the data >1 year: https://www.reddit.com/r/Awair/comments/ys14fk/awair_just_ch...

My biggest complaint with Awair is they only let you download the last month of data easily from the app. That said, they are hacker-friendly otherwise -- for example they offer a local API!

After an accidental emergency call, discovered the cause of this was the "tap to turn on screen" feature -- and my leg was tapping it through my pocket. (On my phone, there was no way to remove the emergency call button on the lock screen.)

Similarly on Windows:

- Open current directory in Explorer:

  explorer .

- Open a file (prompts which program to open it with):

  explorer README.md

Perhaps I was too nice. Most VCs recommended dropping/settling the case, since

(1) it was only a year and many startups go bust anyway;

(2) the best people spend their energy on the future, not the past; and

(3) regardless of the facts, many people are wary of working with someone who sued his business partner personally. (As Trump shows, portraying yourself as the victim of a baseless lawsuit often elicits sympathy, while justifying bringing a lawsuit often requires a careful factual explanation.)

Every communication referenced was in writing. The equity was only one of many very unsavory things that happened, and I eventually had to hire a lawyer. He suggested we sue and took the case on contingency, but by that time the company had minimal assets.