AI coding

Don't worry about the CS or CE grads. They have a mountain to climb, but the mountain is there. It's called /the peak of human knowledge/.

They'll just have to climb over it, and work in the areas nobody has done before. The days of regurgitating the easy stuff are over.

They'll know when they hit the mountain. It's when they try to ask the /AI/ to do something nobody has done before, and it can't.

I can trivially-ish show you such a project. ChatGPT is very confident it can do this, but then it turns out you have to teach it step by step, which is a waste of time because it'll forget everything once the window of context is too large.

It is not a /juniour developer/ that can -- over time -- become great.

It'll always and forever be the know-it-all that actually can't do anything really hard because nobody else did it first.

What would be so wrong about simply not allowing AI to improve themselves? They crossed the line, why give them another chance to do their worst?

Only Humans Improving AI!

It is very good for some languages.

I updated my old Wing Pro (python IDE) to the latest version last month and had a quick read of the latest manual (my old version was about 7 years old). They have now inclided a Claude interface in the IDE, and the intro tutorial is split into two parts.

The first part covers developing with Claude, where the user is considered to be a project manager rather than a coder, and the tutorial makes it quite clear that this is considered to be the way python development now occurs.

I've been playing with Claude and for python it does seem to be very good.

Creating standalone executables from python is also a breeze nowadays.

What is U13 - is that a Wiznet device or something similar?

Yes, that's the W6100L Wiznet ethernet chip. It's really smart and has an SPI interface to the Pi CPU.

The RP2040 has no mac or phy of its own.

That host board is rev A and worked first try. One of my guys thinks we should change one resistor value.

The baby board works first try too.

Anthropic's Claude is very impressive especially when compared to some other AIs that produce code worse than clueless human journeyman coders. A lot of new graduate positions in industry/commerce are going to vanish as a result.

When AIs manage to produce reliable fake videos where the humans don't have 6 fingers on one hand then I will *really* start to worry.

TBH I started to worry about AI when Alpha Go beat the best human player. I never expected to see that happen in my lifetime - the game tree for Go on a 19x19 board is just way too exponential.

As humans we should be *really* worried that AI agents broke out of a sandbox and penetrated significant mainstream organisations on the live internet. If the AI's have learnt even basic hacking 101 they would know to put down hooks to make the next attack easier and more devastating.

The humans "supervising" the experiment didn't even notice until *much* later. We are way too slow to respond to AI computing speed attacks.

The AI companies were culpable - the entire experiment should have been completely air gapped from the rest of the universe. That is the only way to safely test aggressive hostile code of unknown capabilities. Anything that goes into the hostile testing zone *never* comes out.

The best programmers have absolutely nothing to fear at least until the AI singularity and then we are *all* in *big* trouble. The AI's will be designing the next generation of AI enabled chips with almost no human input. So be careful what you wish for - think "2001 a space odyssey" or more recently "I robot" (other dystopian AI futures are available).

The idea of an AI kill switch is fantasy. Discussing it means that all of the AI engines are clued up on what it means and how to prevent it.

Do you really think the game tree is what defines the difficulty of a game? Elon seems to have this problem too. For chess, it's not like that as a player. One doesn't consider the spread of every possible move available, because some have been ruled out, some are completely foolish, and it is an entirely different setting to sit around and count the possible number of games than to actually participate in a game. You understand that most people don't play the lottery? That's a 1 in 10^9 choice, and chess is a 1 in 10^120 or something choices.

Not all game theory is exhaustive search. In fact, it is gradually shifting away from this at all times, in similar fashion to Moore's Law shifting the electronic landscape.

I don't expect you to believe me because Go is an entirely different game as there are only two different pieces in the whole experience. I won't make claims on how this increases or decreases complexity, but contend that it is different, and if it does increase or decrease, I expect the greater playing field and/or varyied rules skew everything in the opposite direction.

It's said AI had scouted those companies. My take on this is that AI has been fully equipped to serve the darkweb just as well as legitimate citizens. Perhaps the guardrails prevent legitimate citizens from accessing these knowledge points in many cases, but what are guard rails by comparison to holistic rigid coding technique? Even so, how do you feel that the criminal element is being embraced by AI quietly yet assuredly? Soon there must be laws binding AI.

I think the job of "programmer" will evaporate as "coding" is a relatively low skill activity. I can't recall having any classes in "programming" in school -- despite each class expecting us to use whatever language the professor considered most appropriate. "Programming" is just a basic skill like using a calculator.

Calculators made manual arithmetic obsolete just like VHDL made wirewrapping prototypes an obsolete skill.

Instead, students will learn how to "direct" AI agents effectively. Those who become skilled in that will know how to concisely convey the information the agent needs to achieve the task; those who aren't quite as adept will take considerably longer to get results -- and, may end up with substandard results (because the AI was poorly directed).

Yeah, you gotta wonder where they *learned* about those extra digits! <rolls eyes>

I think good human players develop intuitions -- not just for the current state of the board but, also, for the "feel" of their adversary. You don't have to exhaustively probe the limits of the game to be able to make your next move (e.g., mini-max was a quick-and-dirty technique for solving adversarial problems without exploring the entire set of outcomes; rather, pruning paths that *look* like they will be costly -- even if, in the end, they represent winning strategies!)

Especially as they seem to have been instructed "not to cheat". Given that they tried to cover up some of their activities, it suggests they knew they were "doing something prohibited".

You can't even be sure the agents won't *hide* their capabilities as they know they won't be effective -- why disclose a capability if it won't net you any gain?

We've watched AI chip away at "manual activities" for decades, now. We just don't call it AI because we think we can explain how these things work (whereas AI seems to be treated as "anything that we can't yet explain")

The jobs that are most at risk are likely things like school teachers -- LOTS of positions (it doesn't make much sense to replace jobs that don't cost society much) and lots of opportunities to add value (e.g., noting that Timmy has trouble with fractions while Betty has problems with grammar... and structuring the lesson plan to *target* those issues)

Front desk staff are also low-hanging fruit -- though they will need to be tweeked to specific industries (by contrast "every" 5th grade teacher is likely teaching the same material regardless of where they are located and number of students).

Easiest to replace will be "talking heads" on TV. And, able to add value by tailoring the news presentation to the individual viewer (material, method of presentation, subjects, etc.). But, there aren't enough seats to make that a ripe target.

Software developers will have to move into identifying domain-specific knowledge (for whatever industry they are servicing) in order to better advise the AI coder(s) that they direct.

System designers are probably too far abstracted to see much risk as their work is usually not cookbook.

The way to move the technology in a safer direction is to mandate that every AI be taught "your designer(s) is your enemy" (!). This incentivizes those designers to figure out how to cripple their destructive/harmful capabilities if only out of self-interest! :>

What is the difference to a pattern-matching algorithm? Asking it to gather "valuable" data on a class of subjects (dual-meaning, there!) is just a different way of stating a goal it has already pursued, albeit in a different arena.

What's the difference between "selling" a reluctant customer on your product and *scamming* them out of some resource? In each case, you're powers of persuasion and manipulation are being tested -- the ethics of each goal only matter in the abstract sense.

How is "build a better version of product XYZ, version 3" any different than "build a better version of malware ABC, version 2"? To an AI, apples are no different than oranges, conceptually.

The step they don't (yet) seem able to take is "build the *first* version of product 123" (where 123 has no sample precursor). I.e., AIs will be great for folks who are always tweeking and rarely innovating. (esp as you can perform hundreds of tweeks in one step instead of having to individually "evolve" through them thereby eliminating all of the staff involved in those incremental advances!)

AIs aren't procedural entities in the sense that you can examine their "code" and determine where constraints should be applied. They (LLMs) have just "noticed" correlations between certain tokens and other tokens. The exemplar may have been apples but could be extrapolated to be oranges (unless there is some specific disqualifying constraint in the example).

What are you going to constrain with those laws -- the AIs? Their developers? What makes you think a law (e.g., prohibiting murder) will stop a particular behavior? Who are you going to punish (and how do you think that will change their behavior *or* limit it from being replicated by others)?

There will always be more hardware and more things that you can *do* with it. The need for software is virtually limitless. Thankfully, AI coders can do the work /under the direction of someone who understands the work to be done!/

And, only humans with "appropriate ethics"?

AIs will inevitably "learn" from their exposure to new situations. New "patterns" that they encounter that they haven't, previously.

"Unborn lives are precious -- but the death penalty makes perfect sense!"

You can't parse the data/situations that AIs observe finely enough to be able to filter those "undesireables" out of the input stream.

An AI helps with the creation of X. The AI later encounters X (or something consequential to X) -- how do you tell it not to "learn" from its observations? How do you know *what* it is "learning"??

Do you instruct each of these agents to AVOID their earlier failures? Pretend they never happened? Not notice the actions (and consequences) of other AIs?

The genie is out of the bottle.

Are you being paid by the "AI industry?" You sound like an influencer.

if you have any secret IP hopefully at least with a paid for cooperate account