I know this has been done in stuff like Minecraft, but do people just try to build logic gates in all video games lol

this is my new favorite image on the internet right now

The Holy Father is speaking about this clearly and people need to listen. Technology needs to serve humanity, not become the means of disenfranchising the masses.

While this is all about Opus 4.8, the real headline here is that they're saying that Claude Mythos is going to be coming out in the coming weeks. You know, the model that was so dangerously good, they couldn't release it. You could only have select companies use it to upgrade their software vulnerabilities, yet even gpt-5.5 has demonstrated really insane abilities at coding and finding security vulnerabilities. Definitely interesting that Mythos originally was a leak, even though it seems to have been a major marketing lever for them. And I'm pretty sure they originally were saying that Mythos is going to cost 5 times as much as Opus. That's insane. Opus on Cursors benchmarks already costs like $11 per task for Opus 4.7, while GPT 5.5 is basically comparable, and it's only like 4-some dollars per task. You're telling me that Mythos now is going to be like $50 per task? Like, no one's going to use it at a consumer level for that. And even the enterprise use cases would have to be pretty compelling. I'm sure there'll be a much to do about it when it comes out, so it'll be interesting to see how this all plays out.

I, thankfully, am already seeing this happen in industry. I'm seeing, you know, various different teams that are starting to adopt different agentic workflows, you know, primarily using coding tools, but also starting to look at it at the enterprise level. And, you know, I really think it's the practitioners that are using this stuff that can really help with the evals the most because they are actually applying the use of the models, which I think is where, you know, pure AI researchers lack understanding how is this really being used, you know, in the world and in industry. So, you know, really excited to see, you know, this kind of stuff because this is really what helps us get the best productivity and quality out of these models as we possibly can.

I can't believe the vacuum in open weight models in the US right now. There has to be someone cooking right now that will open source model weights and build a moat around surrounding pieces. Who's doing this right now I should know about?

With agentic coding tools becoming unsubsidized for enterprise usage, and surely consumer plans soon to follow, how are businesses and individuals going to react? I personally have been crushing features that are an estimated $50-$100 each in token usage if billed at API rates for Claude Opus 4.7 Max. And if I was focusing all day in this it could easily be 4-8 of that, $200-$800 daily. That becomes $50k to $200k annually for that work, which no business is going to sign up for at scale across their workforce. The most obvious lever is to switch to GPT-5.5 which is half the cost for essentially the same functionality. What about when OpenAI jacks up their prices? And at some point what you can realistically achieve starts falling off if you don't have SOTA frontier models behind the workflows. Where does that leave us? Local for individuals seems like the obvious path for serious consumers that dont want to get locked into absurd prices, but I'm hesitant to see enterprises do that at any necessary level. Do we just need really sophisticated harness engineering for this to work with more economic models in larger corporations? Everyone in business wants AI but they're in for a rude awakening as the cost of doing so starts to materalize. Anthropic and OpenAI IPOs will likely push this over the edge. And Flash 3.5 won't save us. Who does that leave?

I think a lot of people talking about this are missing a point here. They're saying that the software itself is useless or that it's not helpful, but I mean look at what it's raised. There's clearly money that is being made and going into it. I think what people are really missing is the video just slapped so hard. Like it almost doesn't even matter what the product is in terms of it going viral.

AI can't be you when it comes to taste. It's often said you need to determine system, architecture, and most importantly taste. We've already seen the first two begin to get eaten up by what agents can do with strong harness engineering, but obviously it still requires significant upfront guidance and direction as well. But it's why AI generated posts and replies are so bad. AI has no taste. Its not that it can't have opinions, but they're so inauthentic, just a averaged out aggregate of the human corpus of knowledge that they have been trained on. They have no skin in the game, it's never a real raw take or having sufficient character for opinionated takes. It feels so safe and calculated, which is boring, ergo, bad content. The real move is to use AI to amplify and increase the throughput of your own voice (literally, use voice dictation) and you will be able to generate so much textual content at speed with a solid model running the transcription. Actually speaking into your posts and replies too gives social media a fresh dynamic that makes it feel more real. If you haven't tried it, I highly recommend doing it. Voice to text mass communication is great.

Can someone please make an agent native version of Excel? I'm so sick of using this spreadsheet tool that has not gotten basic things that have been problems with it updated for over a decade or longer. We are long overdue for there to be a new ubiquitous data tool that technical and business users can use alike. It is far past due. Someone please make this.

An interesting dynamic playing out in industry for agentic work is if it should be a centralized versus a federated model. I think probably there's a little bit of both happening in a lot of companies, but it kind of boils down to a bit of a political struggle between individual employees and their teams and departments against a more centralized unit that maybe kind of better represents the entire business. Because individuals can use agentic tools to be able to increase their productivity and do work, but then you could also conceive of it in a central model of well you don't need all those people you just have less people doing you know more of that work, but then at some point at scale you're not going to be able to have the domain knowledge that those individuals have, but with all of this you know it comes down to who who's getting access to these tools and approvals and who's actually providing value. I think that'll determine really how this plays out at individual companies is how you know individuals working within these federated and central spaces are actually able to execute value for their companies.

Saw a post earlier today about a dude making money selling his GPU compute on online platforms and using solar to help offset the electricity costs and it's really got me thinking about I want to get a solar panel farm to help power GPU clusters I buy in the future. You know everyone's talking about there's gonna become a compute crunch and you know the output you're getting from frontier models is becoming less and less for the same cost it's getting more expensive to have the same but I think also with all these data centers going up we're gonna possibly start having rolling brownouts that happen. So I think you can just be able to have electricity for myself and my family and then for all this GPU compute we're gonna want to secure that with some solar energy if possible seems like a good hedge to bet on for the future of what's coming.

After three attempts to try to start migrating my agentic workflows over to using Emdash I think I'm finally having some success. As someone who had never used work trees before it was very hard kind of shifting mentally to using them. And I think another friction point was I have a lot of chat sessions I'll use with coding agents that are more operational rather than developing or building something. And so those don't even need work trees at all. I think once those two things clicked and I realized how I could actually keep my main branches clean across my projects now I'm really starting to see how much more efficient it is having all my projects in one place. So I'm really looking forward to continuing to work with this tool. Will making using other harnesses so much easier too.

First PC build I've ever done, 2x RTX 5070s in a Fractal Torrent case for local inference, and I've been using Claude to help every step of the way. Used to be you had to watch videos and Google your way out of being stuck, with this I can just endlessly pass images to AI. It's not that it's been perfect, I've had to catch mistakes it's made multiple times, but a lot of it's common sense stuff. Day one was getting the case taken apart, PSU in, cords plugged, and starting on the motherboard. Day two was finishing the motherboard and having my first thermal paste moment, looking at Reddit memes about how much is too much or too little and having Claude try to explain the memes to me, which was something. Day three I put the motherboard in the case and got the GPUs in. Was a bit sketchy, the first orientation had zero clearance between the two GPUs, and it took my wife coming in to basically say why don't you just switch them around. Did that, solid clearance between them, it all fit. Booted to BIOS for the first time, everything's working, CPU not running super hot at idle despite my nervousness about the thermal paste. Next is flashing Ubuntu and seeing what I can actually run on this thing. Before any serious inference though I have to get some electrical work done, otherwise I'm tripping breakers, so that's fun too. I feel like I should really write an article about all of this.

New paper, The Memory Curse, how expanded recall erodes cooperative intent in LLM agents. Tested 7 models across 4 social dilemma games and found expanded memory degraded cooperation in 18 of 28 settings. Kind of verifies something I've already seen. Even back end of 2024 and beginning of 2025 you could sense it when a chat went off the rails, once it had accumulated too much context, it was basically impossible to steer it back. A lot of people conflate context engineering with "more context is better." It's like no, the right context is what's better. You have to aggressively manage the context window, and I think that's going to be big in the coming year or so, more actively managing these context windows in a way that actually makes agents smarter for it. It's not anti memory, you just can't dump everything into context, there has to be an architecture.

An NYT reporter printed an AI summary as a direct quote, retraction issued after publication. Honestly I'm not too surprised. I see it in the workplace more and more too, you know, people have always wanted to cut corners, and now agents are accelerating one's ability to cut corners. These tools will accelerate output but that doesn't mean the output is right. You have to constrain what they're doing and have validation steps to make sure this kind of thing is correct. You could even have more agentic solutions to prevent and correct it. But everyone's interested in how do I abstract away the layers of work I no longer want to do, and verification is probably one of them, you know? At the end of the day if no human ever looks at it, it can still be wrong. These agents lack common sense.

GrapheneOS thread making the rounds on HN about hardware attestation requirements functioning as an anticompetitive moat dressed up as security. Yeah, I mean, in general I am not a fan of hardware attestation. If it's your device, you should have the ability to have full control over it. Something that's phoning home to a remote server to make sure you're only running certain approved firmware and software is completely antithetical to what most people who own their own computer would want, if they were truly aware of and in control of how their data and privacy was being maintained. I think it really goes to show that businesses in the tech field will do anything to be anticompetitive. They'll dress it up in whatever outfit and give it whatever flavor they think is going to stick, especially if they can use regulatory pressure to make it happen. At the end of the day they're just trying to make it so other people can't compete against them. Or even if it starts out with noble purposes, that's inevitably what it descends into. I've been wanting to make the switch onto GrapheneOS for a while now, I just need to get the right hardware and then make the change whenever I can. Stuff like this kind of makes me want to do it sooner rather than later.

Obsidian community plugin got weaponized to deploy a RAT, with attackers approaching targets through LinkedIn and Telegram posing as investors. I mean, you really wouldn't think that going local-first with your own Obsidian vault would be much of an attack surface, but here we are. The actual mechanism though is just the same tried and true social engineering that's always existed, you know, getting someone to install or accept something from a stranger. If you're doing your own stuff locally, the threat surface is yours to defend, rather than a threat surface that's a shared responsibility between you and a cloud vendor. Both have their tradeoffs. At the end of the day it's still the same rule, don't accept plugins, don't accept apps, don't accept files from people you haven't met and don't trust, because you don't know who's a bad actor.