sporkland

Unfortunately for Ramses and Erika the description tapped into something deeper, a symbol that is all too actually present in our society of the infantile male whose every whim is taken care of.

I myself confused the person for the symbol and owe him an apology. But also eff all the child men walking around always taking and never giving. Be more like Ramses.

I've subjectively detected this in previous codex releases where the 2 days before release of a new model the agent went from great to me pulling my hair out yelling at it. I think it's just a win-win for them. They need to ramp up basic capacity and usage on the new model, what better way to free up capacity than to reduce the effort with the current gen. There's not deep detectors that they've regressed (especially before this guy showed solid data).

> Seeing the hype online feels both validating and deeply frustrating. I worked on this literally one year back in March 2025. I spent months of hard work, sweat, and sleepless nights building it

While I have deep empathy for what the author must be feeling. It feels tone deaf to me that as someone working in the model industry, likely using data sets scraped from other folks work, he doesn't see the hypocrisy in this.

We've hit an interesting question in our society where a lot of us are getting tremendous value out of the uncredited and uncompensated use of people's hard work and sleepness nights. It seems naive of the author or anyone at this point to believe that the same won't happen to them.

The question is whether we want this to be how our society operates? I can see a number of pros and cons and I know it's not a straight line from they trained on the works to derived works. But it's hard to imagine as humans how we will sustain broad motivation over a long period if we keep allowing this to happen.

As an outsider in a hyper competitive commercial space like mobile os (apple, android) which part makes you think it's active obstruction and not just moving fast and trying to keep up, get ahead, etc and they want the open source side but it's a lower priority?

Since S3 supports regionally and zonally locked buckets it seems like any promise around this would have to except those buckets where that was configured.

Hacking is also interesting in that it has a clear success/failure outcome so it's easier to RLVR from similar to math problems. Whereas clean code, good architecture, good product taste are reliant on RLHF.

Article brought to you by your local AI agent.

We had one of the most damaging lab leaks in history with COVID and the wuhan lab. And we couldn't even get the facts straight.

I think this is largely a similar thing. The labs should be running at certain levels of containment given the vitality/risk of the organism under study. Hopefully they get there for all our sakes.

But the revealed preference of society at this point is the damage is worth the benefits both in wuhan and with AI. Unfortunately with some of these "substances" it could eventually prove lethal.

Having tried recently to build an app using POI (restaurants) information derived from OSM I've been left both simultaneously in love with the idea of OSM and sad by the reality of it.

The idea that we all have an open map free to use by developers and users not controlled by a rent collecting corporate entity should be wonderful resource for humanity.

The practicalities are often unfortunately far from wonderful:

1. A lot of the POI data in OSM is often either out of date, wrong or incomplete, including names, web addresses, phone numbers, closures, etc.

2. They often don't have addresses as you're supposed to resolve the address from the location, but if you have multi-story buildings this gets problematic.

Part of this I believe is due to the community requirement that this data come from folks walking around vs from providers that say restaurant owners use (square, toast, etc) to manage their online presences.

Google Maps is just dominant in terms of gravity with folks are throwing all their photos and reviews and edits on there. But if google maps had to build google maps on the Places API google itself would be out of business the prices and terms are so out of control.

But even in the more open space you've got:

1. Overture trying to to make OSM actually usable by combining multiple data sources and doing validation. But overture is up to 1 month or more out of date.

2. MapBox and others are trying to make money selling improved POI data on top of OSM derived data with their own sources added.

I ended up building a system that combined OSM (for freshness) plus Overture for reliability plus validation, but no data duplication from google maps to filter bad data coming from OSM and Overture.

If you've ever ridden a bike next to an autonomous car on its worse day and most meatbag drivers on their best day you'd know it's not just life saving... the amount of near misses, stress and asshole moves drops precipitously as well, at least once per trip from my experience, but usually more.

This is way older than 2025, no?

... and? People don't build programs using nand gates that often. I'm not gonna miss people fetishizing missing the forest for the trees in their professional lives instead focusing on other skills that produced real value.

You'll still need to learn stuff and learn your craft but if you don't become an expert at programming switch statements so much the better or the next garbage web framework.

I'm with you on the pros and cons of centralization. But what makes you think if linear became the unified platform that it wouldn't just become like GitHub? See jira.

Given the history here, No matter which tool becomes the death star of developer workflow it always trends towards complexity and the challenges that come with it. Hosting your own doesn't become a panacea as most IT/ops teams also struggle with uptime and upgrades (increasingly relevant).

The only way out IMHO as customers picking two of these goliaths of sending a clear signal to them that we're happy to pay them handsomely for a stable and less sprawling product. Or a hosted open source version with a different cost structure that doesn't drive these wild swings at feature bloat and the inherent instability that comes with that chase.

In college they made us study the case of therac-25 so we understood the dangers of creating too many errors in our software.

https://en.wikipedia.org/wiki/Therac-25

The airlines are certainly bad actors, the problem on the other side is there are a massive number of people that are willing to waste massive amounts of time to get undeserved freebies if the boundary is too permissive. I agree with your underlying sentiment that the needle is too strongly on the side of the airlines at the moment, but I think if they shifted towards a more healthy stance they'd quickly be swamped by these abusers.

I'm curious what the game theoretic optimal here is.

For me the music on the Monkey Island CD-ROM remaster was a real jaw dropper.

I'm not sure why this is surprising.

In the US we try to maintain inflation at a target rate of 2%. For commodities where the costs aren't able to keep pace with inflation you'd either expect prices to double every 35 years or for the product to be half-ish the size (assuming labor costs scale by size). It's a 99 cent candy bar so...

I think inflation rate and costs of lots of things due to oil has jumped so much in the last few years that both the costs have gotten mouth watering and sizes have gotten noticeably smaller. But I prefer either of those to leveraging even more low cost filler, which I assume is the other option.

I honestly think in the US portion sizes tend to be larger than I need here so I'd honestly prefer if they lowered prices and dropped sizes to something reasonable.

A few reasons I'd say:

1. Our incentive structure in software has always been to a large extent driven by more features [1]. Agent driven development has poured gasoline on that fire.

2. Despite all the effort we put into quality assurance by engineering teams, it's still hard to full de-risk releases across a broad set of configurations use cases, etc. Especially as most agents and even the test hierarchy itself tends to favor unit or module tests and these issues tend to pop up at integration points or hard to test non functional requirements.

So I think what we're seeing is people are rapidly adopting agent lead development due to #1 but the risk reduction pipelines haven't caught up with #2. I think they eventually will but we're definitely in for a rough road for a bit.

[1] https://www.joelonsoftware.com/2006/12/09/simplicity/

People keep referring to the output of neural nets as the "average of the training data" but imho the deep value in neural nets and the interesting philosophical part is they are massive conditional probability engines whose results can change based on the variables you feed in.

While I appreciate how deservedly hated all of our politicians are, isn't it such a sad state of affairs that we can't find trustworthy folks to lead these critical institutions?

> There's so much good stuff in this post.

The post starts off so cynical. I know a few people at anthropic and oai and the simplest explanation also matches my observations that they actually believe what they say. That agents will be doing the bulk of the programming in the not so distant future. They believe they themselves will be out of jobs at that point.

They aren't managing some message and trying to teach the anti AI folks a lesson.

I just made a crappy Microsoft account that I don't check or tie to anything. I'm sure it will backfire at some point.

Was hoping to understand if anyone has done any research on it. I could only really find one research paper on the topic [1], but it seems less focused on this specific issue.

[1] https://arxiv.org/abs/2307.09009

I walked into the Mint one night because I liked to grab a bad cocktail and some good karaoke.. some guy got on stage and belted out a beautiful version of Rocket Man. I looked closed and it was Drew.

Guy has pipes.

Oh his software was pretty good too.

Not to discredit anything that was said in any particular blog post.

Folks also need to remember that a lot of blog posts are written by engineers or managers that have their own agendas and careers and often external blog posts can be a form of self marketing or idea marketing that an engineer or director has been pushing internally.

I have no idea if this happened in mozilla's case but the person that wrote it seemed to talk about the their own internal harness / fuzz testing framework quite a bit, and I imagine it was probably a big part of that person's scope / accomplishments and will probably show up at their end of year review and on their resume.

I think as a user I have 2 modes: 1. Q&A mode where it's basically Google search by voice. 2. I'm trying to process an idea I have with an LLM buddy.

My desires are pretty different in the two scenarios. Q&A mode if it's not quick to respond I'll think something is wrong with my phone.

Deep think mode I'm honestly kind of pissed off at how fast it tries to respond. I want it to slow down and give me a chance to process and use extra compute on its side (including newer models) so it doesn't just spew low thought bullshit at me.

It seems like the system could detect which of these two modes was happening and adapt, including protocol.

I haven't tried the voice mode since the new model updates, maybe it's gotten better.

Counter to everything I just said though and germain to the topic at hand, when I'm in q&a mode that's probably the worst time for it to drop audio as it changes the query significantly. vs when I'm talking at it for 2 minutes it could probably throw half away.

I was doubting this for a minute as I wondered with a significantly biased coin towards the head side would you be more likely to get HT. With probability problems like Monty Hall I like to think about extreme cases like say it's 99 heads to every 1 tails. You'd expect HT 0.99% of the time. Ditto TH.

Is there any current research on as agents w/tools start dominating LLM use, if making making models smaller / less single-shot, more like efficient engines that can process a lot of context, and feeding a lot more into context windows is going to be more of a path forward vs trying to memory the world?

Like smaller models that show effectiveness on problems with verifiable rewards when run in a loop with external grounding context?

I went through college around 2000 in compsci and I remember even back then noticing that there were about 10% of us that cared. The other 90% were just in it for money/prestige. The dotbomb cleared them out, but post 2009 you could see the folks creeping back in due to Facebook and co.

We can agree scalpers are net negative.

And I like your ideas but I don't see why the venues and artists don't want to capture more of what people are willing to pay enabled by what the parent comment suggested.

I wonder if in your system it actually attracts fans or just people that have the time to wait for tickets.