sakoht

People say “agents not RAG”, but one framing is that this describes RAG where the database is a file system and bash is the query language (with other cli tools installed it can use, including curl, jq, grep). With writing its own notes on the filesystem structure and maintaining them as a way to “index the database” It is still using code to selectively grab the chunks of data it needs rather than than putting everything in context. It’s just better RAG?

If you like Claude Code or Codex try Warp. It makes both look like a pale shadow of it. And way way cheaper monthly bill.

Remember that for the vent majority of human history, most people have not felt like they were truly in control of their lives. If it was not God, it was the king. Remember that nearly every animal except humans dies by being eaten. We and our pets are nearly the only creatures that die in their sleep. To suggest that the world is supposed to “work” in a way that leaves us feeling un-anxious is a modern sense of entitlement. For most animals, and most humans the only thing you were entitled to was to scrape by desperately until you finally, eventually, failed and died.

Our institutions were more stable before because people were more dedicated to maintaining them. They were created because overt pain or death was the alternative. But then they began to protect us from discomfort. Then from fear of discomfort. And for discomfort that exists in the mind. This was bound to come apart around the time some people asked to be afraid simply said “meh”

Will this go badly? Certainly from the perspective of whether people feel safe and content. But we have had 10,000 years of self generated strive from humans. It is our fuel.

The article suggests a possible model where the schizophrenia is an extreme linear progression. But the inability to find a culprit genes suggests something more complex. It is possible that there are is a group of genes that all have variations that confers benefit, but when those variations are all together negative effects occur. That makes the positive variants overall beneficial, and keeps them in the gene pool. This is why it is dangerous to presume that when we correlate genetic variants with disease and then presume they should go away. In fact, nearly any inherited disease that has survived may be conferring value to other individual when in proper partnership with other genetic profiles.

The odd thing is that among identical twins when one has it the other has it roughly 50% of the time. Which suggests that having the generic basis still requires something external, and it is common but not too common.

If Slack wants me to use their AI to mine my company’s conversations, they need to provide a higher quality product at a better price. Not forcibly take ownership of my conversations.

I’ve loved Slack for a decade, but today I have to plan to exit.

I would argue. They don’t have the right to change these terms on prior conversations. Only new ones.. Because I have a lad slack to hold my companies, important conversations under the belief that I would have access to them later.

I’m sure there is fine print somewhere that says this is not really legally true.

Currently most websites have a snip of code that posts to Google when a visitor arrives. The cookie there is linked to an identity and later used to push ads, and metrics for the site user. It seems like this is the thing to make broadly available. Anyone with $ can index the web. But nobody can get every websites to update to post to 2+ destinations. Multiple companies should be able to listen to that feed, and then be responsible for tracking and indexing on their own.

Reads like a Disney lens on nature. Whether or not wolves have a single alpha in a pack, nature is built around selfishness and violence with certainty, cooperation selectively. Animals watch each other suffer and die without a flinch all the time. And nearly everything dies from being eaten. Any lens on nature that sugar-coats this is projecting human values.

Repeating a note by others, having advice on navigating a large transaction is comforting. Even though I ultimately found the house I bought myself the guidance of my buyer’s agent through the process was invaluable. No regrets on him making good money for his efforts. Also, surprisingly, it didn’t feel odd for him to help me negotiate a lower price while his pay is based on the sale price. His business is based on referrals to my next friend who needs such help.

A similar topic explored in the 2009 small movie Moon w/ Sam Rockwell (and Kevin Spacey as the robot voice). The main character doing labor on the moon figures out he is one of dozens of clones, and the original guy living his happy life back on earth. If you could clone yourself and have the clone take your burdens so your life was easier, is that a fair use of your own body and mind to do so? Clearly the jump to our children says “no”.

Same

> Convincing mimicry is good enough.

For what?

I ask questions and get wrong but “convincing” answers …that is way worse than wrong answers that are obviously wrong.

The functionality looks so smart because it skips reasoning and goes straight to “plausible imitation”. But the latter isn’t actually a path to the former.

If it were reasoning, and merely had errors the problem might-course correct. Even if the errors were huge, and reasoning was poor, it could eventually learn its way out.

It’s not intelligence. It’s a very good simulation of the superficial trapping of intelligence.

It makes $20M a year. That sounds like great code.

I would: 1. Get it in source control without “fixing anything”. 2. Get a clone of the prod server up and running, vs a clown of the db. 3. Put in something to log all of the request/response pairs. 4. Take snapshots of the database at several time points and note where they occur on the log history from number 3.

You now have the raw material to make test cases that verify the system works as it did before, but for bug, when you refactor. If the same set of requests creates the same overall db changes and response messages, you “pass tests”.

First thing to refactor is stochastic code. Make it consistent even if it’s a little slower so you can test.

Once you can refactor, you can do anything. Including a full rewrite but in steps that don’t break it.

If you try to rewrite it from scratch it will probably just never be deployable. But you’d an rewrite it safely in chunks with the above.

The gorgeous thing about these was that, the thing that sold a game was clever playability. Even the games that tried to create an immersive world had to span 4-6 floppy disks, and mostly had to lean on imagination, and your ability to see a fantastic world in colored pixels. I learned to code on a vic-20, then a C128, though the C64 had basically all of the fun software. When the Amiga came out, you had to pay $500 to get a C compiler to build software on it, but it was a dream. When Commodore finally tanked, I looked at the 80286 and Win 3.11, and nearly gave up on computing. Then: Linux. But for all of them, the beauty came because you could "touch the bottom" of the virtual world they offered.

There are a lot of projects that want to serve an under-served space, but lack a novel idea about how to do it. In many ways, even companies like Theranos fell into this trap. They knew it would be great to have the device they promised, but the company wasn't founded around any sort of novel idea about how to accomplish it, just a wish.

Very dad. Hope his family is as well as can be.

It's not impossible to say "Positions vary from $20k to $600k, DOE." for these cases, if you really are hiring your new CTO and also a janitor. The nice thing about having to put _some_ kind of range is that there is some the poster will have to choose whether or not to reveal they _might_ pay well. And there are balanced incentives to go either direction. If you post low, you miss out on talent, and if you post high you have to convince every noob that might be useful they aren't worth $1M/yr. But having to decide that balance is useful.

Regarding interest rates, they are already up. I got a house 4 months ago at 2.85%, and now they are 4.2%. I can't speak to whether investment in tech will drop. Lately, there is so much spare money floating around investors can't find homes for it. Inflation suggests to me _more_ desperation to have the money invested, not less. And the net cause seems to be a steady stream of wealth consolidation allows a smaller subset of the economy to hold more and more wealth.

Fred Brooks' classic "No Silver Bullet" outlines this fundamental principle of software engineering physics. Written in 1986, the paper has its own wikipedia page: https://en.wikipedia.org/wiki/No_Silver_Bullet

"Wouldn't it just be grand if we could write a spec for some piece of code, then let the machines do the rest?"

This is quite possible, if you write the "spec" in a programming language: i.e. write a test case. The problem is that people imagine a formal proof in, for example, English, being simpler or easier to think about than in, say, Python or C. In general this is not true. In general, moment you expect your "spec" (code) to be sufficiently precise for a machine to act upon it, the simplest way to describe it is a programming language.

And if you can invent a better way to describe it, and it is unambiguous enough for a machine to act on it, you've invented a new programming language.

It is sad that, given that the company likely doesn't have a case unless some explicit IP was violated, that threats of court allow money to decide the outcome.

But more sad than that, to me, is that creative energy on both sides isn't doing something more novel. A company doing something an intern can clone. An intern facing a sea of unsolved problems that chooses a space that is covered, as he points out, many times over.

If copying is the highest form of flattery, why flatter your former boss so much?

There is a balancing factor. People tend to overestimate their skill and ability when risks are low, but underestimate them when they are high. An average person is pretty sure he/she can walk on a path that is 2' wide without error ....but if it crosses a lava pit, the instinct changes. Suddenly you don't have confidence, unless you are around lava pits all the time.

Then a second complicating factor comes in: most people interpret egoic/social peril like physical peril. For most of human history being ostracized by the community meant death. So we are willing to believe whatever we can get away with in a social context ...but not more. And the winners of a social egoic competition benefit from this. Competitors self-limit.

There is an illusion that confidence confers skill when what is actually happening is that a diminished ego has impeded basic functioning, and artificial self-estimation counteracts that somewhat.

Even if you aren't in the top 10%, you may be able to solve the problem b/c the average person is impeded by an artificially diminished ego.

Setting aside the evils of the DMCA itself:

Seems clearly that if youtube authorizes downloads, then people shouldn't put things on youtube they don't want downloaded?

The takedown notices should be going to youtube. But of course they want to use youtube to promote themselves.

Presuming if you are 100% sure what "invalid" is: It is possible that being _unable_ to represent a logic error might mean that a logic error is represented as a "valid but incorrect" value, which is even more dangerous.

Take the example of storing a continuous series of date ranges. If I only store the first date of each pair, I can never accidentally have an overlap or gap. But if my code has a logic error that incorrectly calculates a range, being able to represent it could throw an error. If that code error translates to an incorrect break-point instead, I haven't prevented a bug, I've hidden it.

This is a good point. It seems we convict people for crimes they don't know they are committing, and also for thinking they are committing a crime when they aren't.

It seems they should have just had the cops hide and wait until he actually tried to make a move on the undercover "girl" cop. Him showing up out of curiosity, perhaps even concern, is impossible to distinguish actual willingness to hurt a child if the jump on him immediately and arrest him.

The subject of child rape is so terrible, we give a pass to shoddy police work. One would think that a lack of child porn, or overt sexual conversation from the guy would have made the judge think he didn't actually intend to take advantage of her without more evidence.

It's a good question, though one could argue that, if you fail to hurt a person there is no "murder victim", but there is an "attempted murder victim". If you actually make the attempt. If you only "plan" to murder someone, then never go attempt to go through with your plan, I don't believe you have committed a crime. You have the cross the line of taking direct action to initiate the action. Though if you plan with others, "conspiracy" to commit a crime is itself illegal, even if you don't take other action to act on it.

The plot loophole in Minority Report is that they could have just stayed near the victim to protect him/her and waited until the perpetrator actually crossed the line of making an attempt, and arrested the person for attempted murder. Forget the whole central dilemma of the movie.

Sentences that are too short, or too long, are both a byproduct of a lack of latitude on the part of the judge to apply reason and common sense. Our system a judge's "bias", but then end we up with a bureaucracy instead of a human who can see nuance.

Mandatory minimum sentences have the same effect. The prosecutor uses the fear of the sentence to get a guilty plea and negotiates a plea bargain. This means people who should have gotten a fair trial in front of a judge and a jury are scared into pleaing guilty and letting the prosecutor decide how much time they will do by framing the plea offer. We bother to have judges to prevent this stuff, and should use them.

Mandatory minimum sentences have the same effect

So true. Believing we will “finish” needing computer programmers because “the work is done” misses the basic physics that governs this process. And naively believing we are finally there is a regularly resurgent myth.

We have been predicting machines would not need programmers since there were programmers. It is true for a _given_ task at a given complexity level. But overall the demands on, and for, a modern programmer have only gotten higher, because the demand of all business and human activity is to offer more than we might have otherwise.

Just wait until, for an app to differentiate itself in business, you have to create intelligent responses in a variety of augmented reality interfaces, correctly predict human behavior, and interface with the physical environment in a routine and nuanced way. And the companies that can do it well are suddenly dominating the ones who do a sloppy job.

Demand for the things we want _today_ will be met. But progress leads to new demands, and they are more complicated.

Anthropologists estimate that the work week was at 20 hrs at the end of the Stone Age. We have been inventing new problems in the vacuum created by our successes for, literally, millennia.

Google is clearly guilty of neglect because they like to do everything in an automated way ...even if it can’t possibly be done well at this point.

But that is way different than Embrace, Extend, Extinguish from MS. I would not trade, say, Chrome for the world of IE. the tension between Google and Apple is far healthier for developers and consumers.