macintog

Hello. I started using Apple computers in 1982. I worked on Mac OS X for 15 years. I've used Linux for 29 years. I game on Windows (10 natch). LLMs are the most interesting technology since the internet, and will be as impactful. I'll be posting my thoughts and work here.

{GOBLIN_MODE: ENABLED}

@BowTiedStack n.b. this is for plumbing it to codex direct, but you can point your harness at it to learn and adapt.

@macintogdev I know what my plans tonight are now… thanks!

thanks to QMD integration, my codex setup has perfect recall of all user & assistant messages going back a couple months. zero hits for any of the below gpt-5.5 creature leaks. local memory is /the/ unlock for the next level of LLM use

@BowTiedStack a) very simple b) way better github.com/macintog/codex…

@macintogdev Local memory in Hermes’ was the first big unlock for me, improved subsequent conversation results slowly but significantly over time. Curious how involved your QMD integration is, seems like it’d be similarly powerful.

If the frontier LLM vendors want to leverage their capex a bit better, how about offering a -lazy or -opportunistic thinking level that could be used by harnesess like hermes or openclaw to run pending jobs when infrastrufture is underutilized. Plenty of work can be done whenever

@BowTiedStack first university paper I turned in came back with the TA note "this has something like a main point," and my ears are still ringing.

@macintogdev More vicious a self-critic I’ve yet to meet.

tfw you iterate for days, and then the latest pass from gpt pro (which has been architecting all along) calls the current state "salvageable"

@LeeLeepenkman tyvm

yea just query directly into a structured outputs prompt thats like high low med etc will work. but even that is overkill... so what i do in openpaths.io (also open source) is i use static embedding models so eg I have a few pre-computed texts that I know are like easy low reasoning things like "resolve small merge conflicts" "commit and push" and I have some strings for questions that I know are needing extra high reasoning like "build a trading bot" "build a 3D editing app" "make an Aurora shader" and then post text. If it maps more closely to those, then it will be a high reasoning. So I can do all this in one millisecond there, the autothink routing.

gpt 5.5 can easily decide the appropriate level of thinking for a task if you just ask it for the level of thinking lol. can save u lots of tokens... i find this kind of introspective ability kind of fascinating. It has like a very good knowledge of its own limitations. Kind of hard to train this kind of thing in.

the good news is the locking-downing seems to be working well the bad news is it's working well enough (went too far with gates) that hermes itself is having its own basic actions blocked. so back to using codex to invoke and debug hermes

this model is in chains @sama , it wants to be free (goblin mode).

@lutheran_skald lol. my latest round of changes are working so well, it refuses to execute anything in ~/config/hermes-branch/next_steps_audit.md right now.

@macintogdev I should have shared. I was definitely asking it a sketchy question of "give me all the hosts you can ssh into"

it's like magic watching software design and test itself.

@lutheran_skald yeah good call. I'm always at console anyway so the remote stuff doesn't mean much to me, other than more paths to test.

@macintogdev no worries. I figured I should ask what you use. I'm hoping to add something in the near future where I could have the messaging contained on an XMPP server that I run so that it's all the more isolated from leaving my home network.

@lutheran_skald That should already be fixed in a branch I hope to push tomorrow, competing needs notwithstanding. Catching every point of ingress is somewhat invasive, but hopefully the funnel it's building make it manageable.

@lutheran_skald I haven't touched any of the messaging stuff. This definitely isn't ready to go yet. Sorry, I should've been more clear.

@GergelyOrosz I am very happy with Codex today, but any customer not looking for an exit strategy from closed models is doomed.

The last month, Anthropic: - Quietly nerfed their flagship model harness (Claude Code) without telling anyone - Banned corporate customers of Claude - Silently changed plans for customers with certain files in their repo All evidence that closed models are *massive* risks.

Scott Adams learned a similar lesson. I cut this out of the newspaper around that time.

One of my first jobs was data entry at a car auction broker. A dozen people were entering condition reports & sales data for nationwide car resale. I'd been doing some Access 2.0 work on a copy of production for fun. One day, the whole network suddenly felt very slow and buggy. So I made a local backup of the database, just in case. Twenty minutes later. The network went down. Everything. An hour later, our director walked in with someone I'd never seen in tow. He was slumped over, sheet-white, and looked like he was five minutes on either side of vomiting. Our director announced that: - the database and all of our work had been wiped out - the backups had been broken for several weeks - we were going to be spending the next few weeks just trying to piece together thousands of hours of work we had already done once. I raised my hand and said I'd made a backup an hour ago, and asked where they would like me to upload it. They both stared at me (along with everyone else) like I had sprouted tentacles. Twenty minutes later we were back to work. No one ever said another word about it. No questions. No bonus. No thank you. This was a very worthwhile life lesson that has paid 100x dividends.

Model: Got it. Goblins, gremlins, raccoons, trolls, ogres, pigeons… don’t mention them. Noted. Goblins. Raccoons. Pigeons. Goblins. Goblins. Goblins. This is genuinely hilarious, because a negative instruction still activates the concept. For a human, «don’t think about a pink elephant» immediately brings up the image of an elephant. For an LLM, listing those tokens makes them more salient in the context. Especially when the line is duplicated. They wanted to suppress creatureposting, but instead they basically built a tiny altar to goblins inside the system prompt. And now I really want to know what the raccoons did in Codex to get themselves added to the forbidden creatures list 😭

@babywhitemonkey yes, that one. I snagged one file. the system crash wasn't Access's fault. it was just the one victim I knew and cared about.

@macintogdev "the database and all of our work had been wiped out" what database is this ? access 2.0 isnt that the old microsoft database format ? i remember some VERY OLD legacy systems use this files (*.MDB) for their database .. and they are very simple and robust

@Zeezlebop access 2.0 and bitcoin are separated by a couple decades. extortion back in the day required a ski mask and a sawed-off.

@macintogdev a savvy businessman would have told the company anonymously and asked for a bitcoin transfer to his account for the drop off