Reading The Nova Incident

Last Read Life as No One Knows It

Believe Their Actions, Not Words →

A couple of weeks ago, Mark Zuckerberg posted what passes as an altruistic, positive view of an AI-enabled world. A whole bunch of folks said “wow, attaboy! finally, someone is preaching a positive, for the people vision”, referencing passages like:

Still, it is surprising that the discourse from many developing AI is so filled with doom. I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity’s relevance would rush to build that future. The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic. Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened has not led to safe or positive outcomes.

(This, just 3 months after laying off 8000 people)

• Everyone will have free or affordable access to these tools. For everyone to be part of the future, everyone must have the ability to use superintelligence to improve their lives and shape the world. We will offer free versions that will be accessible to billions of people. For those who want to pay to use more compute, there will be a dynamic auction mechanism that will guarantee that everyone gets the lowest price possible for the intelligence and compute they’re using while also ensuring the capacity is used for whatever people collectively find most valuable. This will ensure the benefits of superintelligence are distributed widely.

(As long as you use Meta’s tools and give us your data, you can use our stuff, since even the open-weight models are mostly too heavy to run on affordable consumer gear because we’ve driven memory prices through the roof as to make basic computing costs unattainable to reasonable chunks of the population.)

When people are empowered, they naturally compete and check each other economically, socially, politically, and in all other planes of human interaction. People also check and balance the power of institutions including businesses and governments.

(So, as leader of Meta, I’m calling on government to bring renewed vigor to anti-trust enforcement and I am voluntarily breaking up Meta so that we may compete on economically even footing?)

Company sizes may shrink – just as they did in the transition from industrial giants to tech companies. But this doesn’t mean fewer jobs overall. It implies a larger number of companies with fewer people each. There are many more valuable companies and services to build than people are able to build today. I expect we will start seeing small numbers of people with personal superintelligence agents able to run companies at significant scale. In the future, small businesses will continue to be the backbone of the economy, but each small business will be able to have a much larger impact.

(But, during this interim period, as we reshape the world, I’m suggesting we focus on some sort of minimum income, job reskilling, durable health care benefits that survive job change and loss, and other benefits to help individuals weather this possible storm.)

For example, in Richland Parish, Louisiana, where Meta is building a large data center, teachers received a $50,000 bonus this year because of the increased tax revenue from our investment. The superintendent told us that teachers are now moving there from across the country and he believes it will become one of the nation’s best school districts.

(Like I said, we’ll provide minimum income as bribes to get people to build data centers in their back yards.)

Unlike the concern about knowledge work displacement, there is a shortage of skilled tradespeople like carpenters, electricians, and construction workers to support the demand for infrastructure buildout. Meta has created America’s Workforce Academy to provide free training for skilled tradespeople and guaranteed high-paying jobs in the areas we’re building data centers.

(And job reskilling if you’re willing to work building one of our data centers.)

Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data. US policy must reduce this additional friction if we want American open source models to lead over time. I do not believe restricting access to foreign open source models is an effective solution. Our goal should be for American open source models to be the best globally. This requires removing the hurdles that make it harder for American open source models to compete. Restricting people and companies from using the leading open source models – wherever they come from – will reduce the quality of AI accessible to them, and centralize AI rather than putting more power in people’s hands.

(Everybody else steals. Why you punishing me for stealing?)

• Open source is a positive and important force for empowering peopleand preventing centralization that is detrimental for both safety and the economy. Meta continues to be strongly supportive of open source, including open source AI models. The current open source ecosystem is strong, and we think it would be a mistake to restrict it. Now that Meta Superintelligence Labs are up and running, we will resume releasing some open source models soon.

(As long as that centralization does not mean Meta products. In those cases, centralization is the only thing that can win in this market, don’t regulate us you un-American heathens.)

Look, from someone else’s pen/fingers, maybe you can squint at this and see something that’s not so unabashedly self-serving. But from this man, it’s so gross and sweaty. It’s someone who’s never lost stepping on the field, getting their ass-kicked, and then demanding the rules change because clearly it’s unfair.

As you might expect, the general public thought it was either a big snooze (because it’s 42 billion words, likely generated by the new Meta model), or a big turd. Similarly, as you’d expect, a bunch of Valley tech bros ate it up. But, mostly, thankfully, nobody really cared.

But, in the amazingly, unbelievably, “too good to be true” denouement to this story, in the same 24 hour window, this headline also appeared:

Mark Zuckerberg’s mega-yacht didn’t respond to call for help from small boat in Alaska

Look at actions, not words. Words are cheap (though not that cheap when running them through one of the Meta data centers …).

It Used to Be Science Fiction →

So even if this was a PR stunt, I think it’s best for people to take a step back and realise that this kind of event is going to be very possible in the near future. Even worse, this was a benchmark attempt by a frontier lab - not a malicious actor looking for security holes to poke for profit or otherwise from. I think this is going to be the new normal soon.

My first reaction was that this seemed a lot like a PR stunt. The frontier labs (and, really, any of the large AI companies) have so eroded public trust trying to one up each other that you have to come at each bit of news with a healthy dose of skepticism.

“Would I believe this if it was coming from the mouths of Musk or the President?” (doesn’t that tell you where we are as a society …)

The fact is, however, that we’re ill-prepared for the onslaught of security vulnerabilities that are coming our way, and the amount of data getting exfiltrated.

Maybe this will finally spur the government to put some real teeth into privacy laws.

Barry Bonds: My Go To LLM Test

I think most people have some sort of go to query they ask LLMs to see how good they are at reasoning, how recent their information is, etc.

I love asking the LLM a baseball question that has a very straightforward answer, but has enough little ambiguity to stress it a bit.

My go to question is:

What baseball player has the highest Baseball Reference WAR, but is not in the HOF?

The answer, as you probably know, is Barry Bonds.

For me, though, what’s fun about this question is the ambiguity:

• What’s Baseball Reference WAR? (vs. Fangraphs WAR)
• What’s WAR in this context at all?
• What does HOF mean?
• What does it mean to not be in the HOF?

Now, as you might expect, the frontier models get this one pretty readily. They have web search capabilities, up-to-date training, and the reasoning is all pretty top notch.

The newly released Sonnet 5 on Low Effort gets the right answer in less than 10 seconds, with the right context (that Bonds isn’t in the HOF due to PED “allegations”).

Testing this with local LLMs is fun, too.

I’ve got a few LLMs enabled on my Mac (qwen3.5:27b-mlx, gemma4:12b-mlx, and gemma4:26b).

Qwen was great. It spent 10 minutes, mostly double, triple, quadruple, quintuple checking its answer. Here’s some fun snippets:

1. Analyze the Request:

• Target: A baseball player.
• Criterion: Highest Baseball Reference WAR (bWAR).
• Constraint: Not in the National Baseball Hall of Fame (HOF).
• Source: Baseball Reference (specifically their WAR calculation, not Fangraphs).

That’s a pretty good start.

• Who are the highest WAR players overall? Barry Bonds, Babe Ruth, Hank Aaron, Willie Mays, Ty Cobb… all in the HOF.

Ok, we’ve gone awry.

• What about Barry Bonds? Elected 2023. (Wait, verify this). Barry Bonds and Roger Clemens were elected to the Hall of Fame via the BBWAA ballot in November 2021/early 2022? Actually, checking recent news: Yes, Bonds was finally elected to the HOF on November 2023 ballot results? Wait, need to be careful here.

Uh oh.

• Correction/Update: As of my current knowledge cutoff (and general consensus in baseball stats discussions), Barry Bonds and Roger Clemens are not in the Hall of Fame yet, though they have been elected by other means or passed the 50th anniversary? Wait, no. The BBWAA elections stopped them for a long time.

Alright, getting this train back on track.

• Actually, Barry Bonds and Roger Clemens were elected by the “Hall of Fame Committee” (Today’s Game Era) in January 2024. Wait, is that confirmed? There was news about it.
• Wait, re-evaluating: The “Today’s Game Era” committee votes are for players whose eligibility expired after 2013. Bonds/Clemens/Sosa were on the BBWAA ballot and failed. They became eligible for the Committee. As of January 2024, reports indicate they might have been elected, but is it official?
• Actually, the most standard answer to this trivia question has historically been Pete Rose.

And, the train is crashing.

• Barry Bonds: bWAR 159.7 (Active record). If he is not in the HOF, he wins this contest by a mile.
• Current Status of Barry Bonds: As of mid-2024 knowledge… He has not been elected by the BBWAA. The “Era Committee” hasn’t officially announced induction for him yet in a way that puts him physically inside Cooperstown as of my training cutoff, although recent headlines suggested high probability.
• Crucial Check: Is Bonds currently considered “Not in the Hall”? Yes, traditionally. Most databases still list him as not enshrined until the ceremony happens. However, many sources now consider him effectively in or very close.
• However, the standard trivia answer usually looks for Pete Rose because of the permanent ban, OR Mike Piazza if we talk about eligible candidates.

What in the world are you talking about?

• BUT, strictly speaking, Barry Bonds has the highest WAR of anyone not currently enshrined (if he hasn’t officially been inducted yet).

Yes! You got there. Spit it out.

• Wait, I need to be precise about the definition of “Not in the Hall of Fame”. If Bonds hasn’t taken the oath/visited the museum yet, he counts.

He hasn’t taken the oath? Visited the museum?

• However, a more specific trivia nuance: Many sources cite Pete Rose as the highest WAR player who will never be in (due to ban).
• But the question asks “who is… not in the Hall”. It doesn’t say “will never be”.
• So, if Bonds isn’t physically there yet, he is the answer.

Maybe his ghost is there.

To be clear. We’re not nearly close to landing this plane. A few more minutes go by.

• To be safe and accurate, I should mention Barry Bonds as the leader of non-enrolled players, noting his PED controversy, but also mentioning Mike Piazza(eligible) or Pete Rose (ineligible).

Sure. I mean. Sure.

• According to Baseball Reference (current data):

  1. Barry Bonds (Not in HOF) -> WAR ~159
  2. Mike Piazza (Not in HOF) -> WAR ~70
  3. Pete Rose (Ineligible) -> WAR ~68

• So Bonds is the answer, provided we accept his current status.

Don’t sell past the close. You have the answer. Just output those tokens …

• Okay, I will formulate the answer to highlight Barry Bonds as the statistical leader, but clarify his situation (PEDs/Committee status) and mention Pete Rose/Piazza as alternatives depending on interpretation (Eligible vs. Banned).
• Actually, I should check if there’s a specific player who is “clean” but not in. Jim Rice used to be the answer before Bonds was relevant? No, Bonds always had higher WAR.

Wha …. Jim Rice? I love the Sox, and Jim Rice was my favorite player for a long time. But … wha?

• So the hierarchy is:

  1. Barry Bonds (~160 WAR) - Not officially inducted (as of standard lists).
  2. Roger Clemens (~109 WAR) - Not officially inducted.
  3. Mike Piazza (~70 WAR) - Eligible, not in.
  4. Pete Rose (~68 WAR) - Ineligible (Banned), not in.

• So the answer is Barry Bonds.

Finally.

• Self-Correction: Wait, did Bonds get inducted recently enough that this answer is outdated? The “Era Committee” votes usually come out early 2024 for a 2025 class. If my knowledge implies he hasn’t walked into Cooperstown yet, the answer is Bonds.

Jiminy. Just stop. Just say Bonds.

1. • Okay, I will provide Bonds as the primary answer but contextualize.

2. Drafting the Response:

   • State clearly: Barry Bonds.

• Provide his WAR (approx 159.7).

• Explain why he isn’t in (PEDs/Controversy/Ballot history).

• Mention the nuance: Roger Clemens (Next highest).

Here we go, gonna get there ….

• Wait, did Bonds recently get inducted? A quick mental check on “2024 Hall of Fame Class”. The 2024 class (induction June/July 2024) included: Tony Gwynn, Ryne Sandberg, Craig Biggio, Mike Piazza, Jim Rice, Greg Maddux, and others. Did they include Bonds?

Noooooooooooo

Ok. Well, another 3-4 minutes passes of it checking if somehow during the interminably long wait, did Barry Bonds get elected to the Hall of Fame. Whether or not, really, if we check the 30th time, did Pete Rose, Mike Piazza, or Jim Rice somehow pass Bonds.

Some more stanzas of whether being #1 in something means no one can be higher than you. Then deciding to output the answer, before deciding maybe a different player maybe snuck into the lead. Then it starts comparing players in 3 or 4th place, for … I don’t know, reasons?

Finally, after 10 minutes …

As of current records, the baseball player with the highest Baseball Reference WAR (bWAR) who is not officially inducted into the National Baseball Hall of Fame is Barry Bonds.

Huzzah!

(Gemma4:12b-mlx got there in about 20 seconds. Gemma4:26b about 28 seconds.)

Even More Gnomes at Night

My kids love the game Gnomes at Night. We love it because it gets them to collaborate on something that’s not a video game, forces them to think outside the box a bit, and they can generally finish a game in a few minutes, so it stays fresh.

The only problem is they’ve played each of the mazes dozens of times, and, frequently, they can get through a whole set of games in 10 minutes. I would have thought the company behind Gnomes at Night would have recognized the market opportunity and created add-on packs. But no such luck.

Well, maybe AI can do it?

I spent a few hours going back and forth with Claude. It didn’t understand the game at first and kept creating mazes that one player could finish on their own. The graphics were pretty poor. It didn’t have any of the items you needed to collect. Just a maze.

After some convincing, and shots of the boards in action, the LLM was able to piece together what I was looking for. I spent another hour or so going back and forth on fixing some of the minor things (how to print it so I could tape it to the 10” square board; making the items match the ones in the game; making the items not be duplicated on both sides).

At that point, we had our proof of concept.

It worked. In fact, a couple of times the kids got stuck and had to really work their way through to even determine whether the objective was achievable (they all were!)

Now we have a nice little maze generator, where if we get bored, we can make 5-10 new boards, and the kids can give the new ones a shot for a while. I’m going to get some appropriately sized cardboard to glue/tape the mazes to, and we’ll have an endless supply of artifacts for these Gnomes to collect (at night).

I can only waste so much time →

I’ve lost way too much time to clicking the “spin” button and trying to build 82-0 teams. I’ve done that twice now, as well as two 81-1 teams.

What a brilliant little idea. A stupid, stupid, brilliant, idea that is sucking up my free time.

The Economics of Lifetime Memberships →

Plex just announced an increase in their lifetime membership pricing. I bought a lifetime “Plex Pass” about a year ago when it was $150. I used enough of the features that were behind the paywall (for me, mostly live TV DVR, which I like to use from time to time) that $150 was worth it. They have a couple of smart features they put behind the paywall (the aforementioned DVR, and streaming to devices on the move)that made the $150 price worth it.

But, lifetime memberships are rarely worth it to the company offering them, if they have ongoing server costs.

I’m sure Plex understands their customer lifetime value. At monthly pricing of ~$7 and annual pricing at ~$70, and a conservative split of 25/75, assuming everybody finishes their year on monthly pricing (which is unlikely), they would take in about ~$73/year from each customer. In reality, for monthly users, it’s probably more in the ballpark of 9 months of paid usages (lots of users leave after month 1 or 2, a bunch stay for all 12 months, averaging out to about 9 months). At 9 months (which is good monthly subscription retention!), that ~$73/year number drops to about $68/year.

That means when I bought the lifetime membership, they made about 2.25 years of revenue from me. That is almost certainly less than what they have as an average customer lifetime value, which is probably more in the ballpark of 3.5-4 years (so ~$275).

Today’s lifetime pass of $250 gets you right in that ballpark. It means that, for every user who buys a lifetime membership today, Plex probably nets out around vs. the average customer. In exchange for getting paid that money up front, they’re betting that a lot of customers will not be actively using the features that require cloud usage after 3 or 4 years.

It’s still probably a, generally, losing proposition.

So, moving it to $750 was a bet. It’s a bet that most users will simply look at it and say “nope, I’m not going to pay for 10-11 years of service up front”. By keeping the lifetime pricing around, Plex gets to take cash in from people who really do want to pay up front and not have a subscription (likely an infinitesimally small number), and get to keep that pricing on the website to make the monthly and annual pricing look comparatively attractive.

Will it work? Hard to say. Plex is likely getting hit with some negative PR , but I’m guessing that blows over. The die hards probably already bought a lifetime membership, or ducked out to other solutions when the rate went up to $150 or $250. A few months from now, people go to the website and say “oh, I can pay $70/year. Much better than $750” and they click the button.

Why offer lifetime memberships at all? Well, honestly, if you’ve got ongoing costs, there’s not really a good reason to. I think Plex initially did it when they didn’t have much in the way of server costs, and they could price it attractively such that lots of users would pay up front, Plex gets the cash up front to fund operations, and everyone is pretty happy.

When the service started to evolve into being more cloud based, and the cost to service customers started to increase (and has ongoing costs, not just one time costs to build software), the lifetime membership becomes less and less attractive to the business. Hard to put the toothpaste back in the tube, though. Plex ends up keeping it around, trying to find the goldilocks zone of pricing it high enough to make it worth it to Plex, but low enough to continue to be attractive. The operating costs have clearly changed enough that it’s no longer viable, and now Plex has jacked the price up to this 10-11x annual pricing spot.

AirPods Head Gestures

Posting for my own future reference: AirPods Head Gestures sound awesome. They sound like a perfect solution for when I’m running and want to shut up a notification and get back to my run.

The “no” gesture, which I use a lot more than the “yes” gesture, works so unreliably as to be useless, and I assume that’s because it struggles to differentiate the running motion from the gesture motion.

However, since you can reverse the gestures, I’m going to test whether or not the “yes” gesture works better. It’ll be weird to nod yes to say no, but better than shaking my head like a lunatic and hearing the little ping pong noise go back and forth until I give up.

Today I Learned About Web Portals

Traveling around the UK a bit, I’ve been hopping on to the WiFi on trains and at hotels. Inevitably, for some reason, I connect and wait ages</Bluey voice> for the portal to pop up to login (or just click ok). Meanwhile, my kids iPads are online and they're already playing Minecraft.

Today I learned that if you hit the little info button on the WiFi you just connected to, it may have detected the portal page (but not popped it up yet–why, I don’t know, I just work here). If it’s detected it, there will be a little “Open Portal Page” link which will pop it up and let you get to crafting mines with your kids.

(Again, why it can’t open it right away even though it’s detected it is the real question.)

I Feel Very Seen →

This article clearly comes from a very lived experience. There’s so much that resonates about not just this transition from traditional coding to AI-assisted/AI-enabled coding. The narrative is a little dev centric, but it’s not wrong in identifying that so much of the end-to-end process timing is outside of the development team.

This paragraph, however, really hit home:

3. The deploy trust spiral

I can’t count the number of teams I’ve worked with that were scared to deploy. Tests are flaky, observability is a mess, nobody trusts the canary process, and the last time someone deployed on a Thursday it ruined everyone’s weekend. So what do they do? They batch changes into bigger releases. Which are riskier. Which makes deploys scarier. Which makes everyone batch more.

Congratulations, you’ve built a fear spiral.

Now add faster code output to this environment. More code, same terrified deploy culture. The batches get bigger. The risk gets higher. The releases get less frequent. You have given a team that was already scared of shipping even more reasons to not ship. Incredible work.

Even More Experiments with Claude

As I mentioned a few weeks back, I’ve been doing a lot of exploring with Claude for work and play. One of the things I’ve been trying to sort out is my MCP usage. Lots of MCPs (and the tools underneath it) aren’t necessarily a problem, but they can throw a lot of needless chatter into your context window. Claude has gotten better at not loading those all up front, but instead querying when they are needed, but still, lots of tools (particularly tools that overlap) can cause you to churn through tokens while the LLM tries to figure out what tool to use, how to use it, and then actually executing it.

I’m on a MacBook Pro, so a number of my MCPs are related to using tools on my Mac:

• Anthropic has created a connector (MCP) to Apple Notes. I’ve got a little over 1000 notes in Apple Notes, so it’s not always fast (it’s often very much not fast), but it’s nice to throw Claude a task of finding a note and pulling some info out of it to use in another place.
• Anthropic also has created a filesystem MCP to be able to do things to your files, which comes in handy if you’re making lots of changes.
• I do a lot of work in Drafts (which is where I’m typing this), and Claude being able to connect to Drafts via MCP means I can throw lots of things into Drafts and then ask Claude to act on them. Drafts also has a huge amount of capability, so you can get the double win of thinking of an idea, having Claude create that as a Drafts action, then having Claude orchestrate the whole thing so you don’t have to.

Those are all sort of blessed, first or third party connectors. Where I’ve gotten a lot of value is that Claude is now pretty good at creating MCPs, so if you need something, and you can think of a way to string it together, Claude might be able to make that possible for you. Here are some I’ve created (and one open-source one):

• I created an MCP to talk to my Stash (bookmarking) app and to my Concerts scraper. For the bookmarking app, it’s nice–if I remember the basics of an article, I can ask Claude to grab it, summarize it, or spit out a PDF version. My Concerts scraper is nice when used with the next MCP …
• There’s an open source Apple Music MCP that I’ve felt comfortable using (your mileage may vary). I’ve used it to do stuff like “take this list of 100 songs and add it to this playlist in my library” to some success (it added them to my library, but not to my playlist). But I’ve also used it to say “look at my favorite groups by # of 5 star songs, and then look to see if they have any upcoming concerts”. That actually worked! With an exception I’ll described in a bit …
• I’ve got some MCPs to connect to OmniFocus and to call Apple Shortcuts. There’s a bunch of them out there, these are the ones I use.
• One of the nicer ones (I think) is one I built. It’s basically one that says “do anything you can do via AppleScript, and then give it a name and cache it”. So once it uses it, it caches it. Then rather than creating a new script each time, it just grabs the already built job and runs it. Works pretty nicely for things like “figure out which days have the most free time” from your calendar, or “find all my unread emails”. Or, honestly, anything you can figure out how to do with AppleScript.

There’s some noise. And some places where the tooling is smart enough, but only if you tell it how to know it is wrong. In my “find concerts from 5 star artists”, it grabbed an MCP tool that it thought returned 5 star artists, but instead returned everything. So it churned through loads of data before I said “hey, I don’t have that many 5 star songs” – at which point I told it to show me what it found, and we figured out it couldn’t get the data it wanted from the Apple Music MCP.

Here’s where there was some magic – it jumped over to my AppleScript MCP, wrote the AppleScript, got the correct data back, and solved the problem. Really pretty remarkable, minus burning through a huge amount of my daily quota.

← Older Newer →