badloginagain

If remember correctly, this was put in by Trump first round, set to activate when Biden was in office.

Now Trump second round fixes it, but expires in next (presumably) Democrat administration.

Which is funny because US industrial investment was on a tear pre-tariff as companies near/on-shored at historic rates. Not only were tariff's not needed, they've effectively shut down their intended goal.

I hate the fact that CI peaked with Jenkins. I hate Jenkins, I hate Groovy, but for every company I've worked for there's been a 6-year-uptime Jenkins instance casually holding up the entire company.

There's probably a lesson in there.

Yo I just heard about MidJourney this year.

And this appears to be a local runtime stable diffusion streaming library?

Bruh.

Now that you mention it I have to agree, they've released a ton of stuff.

Actually biggest complaint is they don't continue supporting it half the time, but they OS a lot of things.

Thanks for the link, interesting.

I just don't see Meta's role-- where is their monopoly in the stack?

MSFT seems to have a much clearer compliment. They own the servers OAI run on. They would loooove for OAI competitors to also run in Azure.

Where does Meta make its money re: AI?

Meta makes $ from ads. Targeted ads for consumers. Is the play more better targeting with AI somehow?

So Microsoft's definition of winning is being the host for AI inference products/services. Startups make useful AI products, MSFT collects tax from them and build ever more data centers.

I haven't thought too critically yet about Meta's strategy here, but I'd like to give it a shot now:

* The release/leak of Llama earlier this year shifted the battleground. Open source junkies took it and started optimizing to a point AI researchers thought impossible. (Or were unincentivized to try)

* That optimization push can be seen as an end-run on a Meta competitor being the ultimate tax authority. Just like getting DOOM to run on a calculator, someone will do the same with LLM inference.

Is Meta's hope here that the open source community will fight their FAANG competitors as some kind of proxy?

I can't see the open source community ever trusting Meta, the FOSS crowd knows how to hold a grudge and Meta is antithetical to their core ideals. They'll still use the stuff Meta releases though.

I just don't see a clear path to:

* How Meta AI strategy makes money for Meta

* How Meta AI strategy funnels devs/customers into its Meta-verse

The real danger I see is targeted AI scams against high value targets.

How much money would someone be willing to spend to capture Elon entering his eX-Twitter password?

Arguments:

1. The danger of AI is they confuse humans to trust them as friends instead of as services.

2. The corporations running those services are incentivized to capitalize on that confusion.

3. Government is obligated to regulate the corporations running AI services, not necessarily AI itself.

---

As a counter you could frame point 2 to be:

The corporations running those services are incentivized to make/host competitive AI products.

---

This is from Ben's take from Stratechery: ( https://stratechery.com/2023/openais-misalignment-and-micros... )

1. Capital cost of AI only feasible by FAANG level players.

2. For Microsoft et. al., "winning" means being the defacto host for AI products- own the marketplace AI services are run on.

3. Humans are only going to provide monthly recurring revenue to products that provide value.

---

Jippity is not my friend, it's a tool I use to do knowledge work faster. Google Photos isn't trying to trick me, it's providing a magic eraser so I keep buying Pixel phones.

High inference cost means MSFT charges a high tax through Azure.

That high cost means services running AI inference are going to require a ton of revenue in a highly competitive market.

Value-add services will outcompete scams/low-value services.

Let your ears decide: https://www.youtube.com/watch?v=ghUs-84NAAU

And we're seeing the result in real-time. Stupid shit doers have been replaced with hopefully-less-stupid-shit-doers.

It's a real shame too, because this is a clear loss for the AI Alignment crowd.

I'm on the fence about the whole alignment thing, but at least there is a strong moral compass in the field- especially compared to something like crypto.

I feel we hold up single-observability-solution as the Holy Grail, and I can see the argument for it- one place to understand the health of your services.

But I've also been in terrible vendor lock-in situations, being bent over the barrel because switching to a better solution is so damn expensive.

At least now with OTel you have an open standard that allows you to switch easier, but even then I'd rather have 2 solutions that meet my exact observability requirements than a single solution that does everything OKish.

"Here is an attempt at ethical, non-sexual haikus for and against Kubernetes"

Amazing how far we've come.

We then train the LLM on all code optimized like this.

How "far fetched" this is from click fraud is the central argument, and I can easily see a range of arguments from both sides.

On one hand, distributing and marketing a tool that is explicitly designed to harm trackers can be seen- ie lobbied- as fraud.

On the other, this isn't a targeted campaign of misclicking, this is essentially random noise on a user by user basis.

At scale though, this fucks with a whole lotta peoples paychecks.

Make it a thread on X

US healthcare could accurately be described as the literal worst case scenario.

If I understand correctly, the meat of the argument is "that is a system for every (∀) task, there exists (∃) a setting that gives the correct answer for that one task."

My understanding of this (correct me if I'm wrong) is that the scam is convincing users that GPT-X can do anything with say, the correct prompts.

This argument misses the mark for me. It's not that it solves all the problems, it's that the problems it does solve is economically impactful. Significantly economically impactful in some cases- obvious examples of call centers and first-line customer support.

Which is insane to think about given the hard open-source "engineer first" pivot Satya has taken the company.

It's like Windows team is committed to building the complete antithesis of that vision.

Peter Zeihan's take is that urbanization leads to less children because there's less space, you don't need the free labor kids provides on the farm, and children are very expensive in the city.

This is coupled with the speed of urbanization for countries that industrialized after the second world war- the later you industrialize, the faster that industrialization happens, the more stark the transition to a childless economy.

As mentioned in the article, there is a demographic boon for that industrialized generation. Less money needed for schools, etc, more time your prime working age adults can contribute to the economy.

Except all those countries industrialized around the same generation. That generation is aging out of the workforce and there's nothing to replace them.

Zeihan posits this leads to demographic collapse, and that these countries just simply "go away" because there isn't enough children to keep the country functioning. I'm not sure how much I believe that, but I do know that nobody has a clue how to fix it. Japan has been front and center for this problem and still haven't found a way to reverse the trend.

huge raging asshole

That's not really how he described it.

His point is that the raw model that became GPT-4 would do literally anything it asked you to.

It would write fascist propaganda just as readily as it would offer medical advice. Literally any and all input from the user was fair game.

But it wouldn't just veer from medical advice into fascist propaganda, not unless the user was steering it in that direction.

2010's were definitely all about putting a phone in every humans hands.

I'm still on the fence about Substack.

I'll have people I follow use Substack, the trick being after some amount of time their work becomes premium access only.

On one hand, I suppose paying people for their insights is legitimate.

On the other, I'm not paying for things I only have a passing interest in.

I dunno, maybe if there were Substack credits, where I could unlock a premium article on an ad hoc basis. Kind of like Audibles subscription service.

But that might cannibalize direct subscription.

You could start eating Patreons lunch and have tiered premium levels for creators, incentivizing community building through Substack. Allows for both Substack premium subscription and creator subscriptions.

In theory that sets you up to take on ever bigger platforms like YouTube and Twitch, getting creators to use Substack for their video/steaming content.

Anyways, forgive my rambling, that's the monetization route I'd look into to have Substack be the only platform any creator needs.

But they haven't built any cybertrucks

I forget where, but someone from OpenAI said that that the next GPT is "all about the data".

I wonder if you have pre-training data that is fully factual, high quality, with perfect logic will result in a LLM that doesn't fall into reasoning gaps humans fall into.

Or are these reasoning gaps an innate component of intelligence?

Probably the most exciting thing about AI development is we get to start testing the things that make us uniquely human.

There's a bunch of loosely connected points here, but the author seems to take umbrage with the inability of GPT4 to break its alignment training of referring to itself as an AI language model.

If anything, the author is demonstrating that alignment is going better than the doomsdayers are admitting. The LLM is prevented from offering an opinion on the best episode of Love Boat, despite (presumably) losing the author as a future customer.

Alignment training does come at the cost of performance- the author was not able to achieve their goals of talking to an AI pretending to be a human- but as an AI language model, it also won't teach that person how to build a bomb.

That makes no sense. The one thing chatgpt does _really_ well is setup unit tests, which is the part of unit tests I hate.

My ChatGPT workflow is give requirements -> have it create unit tests -> give it test results until it passes.

Been playing around with a generated-code-only project: https://github.com/JerkyTreats/scrivr/

In that workflow I don't really look closely at the code. In most cases I've found it isn't really necessary.

I'm glad our AI overlords are as bamboozled by python vs. python3 as us lowly humans

If they lost $3.3b, it's what, .9175 : 1 ?

So nominal price should be $.9175 on the dollar presumably.

Add in another couple bips for the uncertainty.

Edit: It would be actually hilarious if USDC failed because the banking system. I doubt it will happen, you'd have to see further contagion from the other banks Circle uses for their dollar reserves. I figure the tipping point for a bank run is quite high?

1) Preparing to leave my job once my spouse finds one. I need a few months off to get into a healthy mindset.

2) Dive into AI, learn how to do small scale LLM end-to-end experiments

3) Pick up my game experiment on Heirarchical Task Networks, based on Maslow's Hierarchy of Needs.