It never had to sound plausible. AI didn't cause MAGA to believe Haitian's were eating people's pets. AI didn't cause them to believe that schools were buying kitty litter for furries. It's a population who has been conditioned over decades to believe the most wild shit as long as it comes from one of their "trusted sources" and anything that is counter to that is the "liberal media". Fox News et al are far more damaging to our society than generative AI is.
I think their use of the word “plausible” was in reference to a comparison to content generation in the pre-LLM era, which was limited to GANs and Markov chains.
It's an incredibly inefficient way to make a game, but you could check out Handmade Hero by Casey Muratory. I'm not speaking at all to the overall quality of the work. But he definitely bootstraps the game from scratch and there is a ton to learn about the low level graphics and input systems.
Day 004 is where he actually starts drawing on the screen. So about five hours in if you skip the C intro. Or you could start writing to the screen in under 30 minutes with some basic libraries like SDL. In my experience most game devs, even those who want to write the code and not live inside Unity are using some OS/Graphics/Input abstraction library.
As much as I prefer Claude, I cannot for an instant believe that OpenAI is receiving less traffic. Maybe they are receiving more traffic that is easier to shift to worst models like the ChatGPT interface which is probably a huge percentage of OpenAI usage. I'm genuinely curious how much load for both OpenAI and Anthropic is split between their chat models and their agent harnesses.
Anthropic has one of the best status pages of any technology I use on a regular basis. Every time I've had an issue, the status page reported an issue. A vastly different experience from Azure or AWS or GCP or honestly most services which pretend to maintain a status page.
Seriously take a look. Compare it to basically any other status page that companies make available. Their results are not flattering with all the downtime and issues. But it's far more transparent than most services I've experienced.
I think it's great that you were able to build all of that! It sounds very useful. I'd recommend asking Claude Code to evaluate your entire setup and try to identify areas where a script running on a schedule could replace parts of this framework. You could probably get 90% of it from simple repeatable scripts which would let you save more tokens for things that actually need LLM "intelligence". But if you're happy with the ROI then it seems to be working quite well for you!
I actually have a /loop running scheduled tasks like syncing customer data from routeboard (my job scheduler) and a flat file (MD and JSON) agentic CRM I have created that keeps tabs on the status of all the jobs, invoicing etc. This happens at 2am every day.
The great things about NanoClaw is that its actually Claude running in an Apple Container on my Mac. I gave up on OpenClaw fairly quickly because it seems like the biggest security regression ever created in the history of human kind.
I have a Max 5x plan and it'm very happy to pay the money TBH, considering that proposals take only 20 minutes to build, and my conversion rate is quite high.
When I send them for the first time to a new real estate agent they love them because it's easier for them to "sell" my service to their client.
Just be careful. By all accounts I don't believe that the subscription plans cover 3rd party harness support. You may be at risk of all of this being shut down. That's one of the reasons I haven't explored other harnesses. The main one being, I can accomplish everything they can from within CC with some extra steps I know how to do compliant with a CC Max Plan Terms of Service.
Yeah I know, I have experimented locally with CC channels - but it's brittle and buggy at the moment. I will give it another go in a few weeks once they iron out the kinks.
If you do run into any troubles you're not sure how to navigate, please feel free to reach out to me. Hell, even before that. I'm happy to collaborate on your LLM experience. But if the worst happens, I can probably get you back and running at that baseline even if Anthropic bans you. And this isn't me trying to worm my way into you paying me. Just a better understanding of what you've implemented versus how I typically use these tools and if I can close the "token" gap.
Never felt a need for it. I can already replicate much of what it does in more sustainable / preferable ways. I don't want agents reacting to things and doing things. I use agents to build reliable scripts which are then automated. I do have data collection points that I use an LLM to evaluate. The last example of this is I built a job polling service using CC. It's just a normal script that hits open APIs to pull job listing results into a SQLite database. I have another report which is run that drops an update of how many new jobs are in the database. If there's enough for me to be interested in, I'll fire up CC and have it parse through the job opportunities for the ones which match the profile we've been building. I've used an agent to literally build and deploy it all and it runs on an automated schedule. It just doesn't do agent shit while I'm not looking.
I could have piped the results of that search into `claude -p` and had it do the analysis "real-time" and only alert me about things I would be interested in. That's closing the loop in a similar way people use OpenClaw for. But I'm just not interested. It adds more failure points and conditions. The automated things should be as simple and predictable as possible. This may change after months or years more of LLM development or even just me refining my working config. But not yet.
I don't use any other harness, but I have a cron that picks up changes in my jsonl every X minutes and writes them to a SQLite database with full text search. I also have instructions in my user level claude.md (applies to all projects) to query that database when I'm asking about previous sessions. That's my primary use case where I want it to grab some specific details from a previous session. I have terrible context discipline and have built some tools to help me recover from just continuing a different task/conversation with the wrong context.
I could search it myself, but haven't needed to. Getting it out of SQLite into some format Cursor understands should be trivial.
Incredibly disingenuous for nuclear power proponents to state that grid storage is expensive. Your entire argument centers around the most expensive power generation available and one of the slowest to build.
Grid batteries are mostly four hours duration because that's the natural charging and discharging rate for current chemistries, and there's no financial case to store energy for longer right now, as in there's no demand for long term storage at all. Four hours of storage is super profitable, eight hours is just unnecessary and nobody pays for it.
Plus, referring to this in terms of "hours" isn't exactly the most informative reference, the more underlying characteristic of storage is if it's meant for daily cycling, meaning that it makes mine every single day from arbitrage, or if it's multi day which is not a super profitable form of arbitrage at the moment. Daily cycling makes a ton of sense with solar. Longer term cycles make more sense paired with wind.
I think this boils down to more discoverability of the entire API. While I'm not a fan of GraphQL necessarily, it does provide the tools for very robust LLM usage typically because of the discoverability and HATEOAS aspects they actually follow compared to most "REST" APIs. I would love if LLMs could learn everything they need to about an API just by following links from its root entry point. That drastically cuts down on the "ingested knowledge" and documentation reads (of the wrong version) it needs to perform. Outdated documentation can often be worse than no documentation if the tool has the capability of helping you "discover" its features and capabilities.
reply