This weekend on The Servitor: A look at how LLMs went from 2023 barely getting web access to today's OpenAI "Operator" and "Agent Mode." These newer modes try to give the Agent's their own "computers." They're a mixed bag. Some people are finding success with […] [Original post on sigmoid.social]

This week on The Servitor: How the advent of agents is snowballing token costs. https://theservitor.com/token-economics-are-serious-ai-business-api-costs-are-out-of-control/ #ai #API #agents

This week on The Servitor: Abi's been watching Black Mirror again. It hits different these days because we are so there. https://theservitor.com/we-re-living-in-black-mirror-episodes-and-nobody-seems-to-care/ #ai #opinion #tv #blackmirror

Got a lot of personality of Claude out of nowhere. Was having it analyze/translate some Korean news for me. I don't like a lot of personality in my AI. Me: "The market briefly touched 92 and has been trading in the low 80s. i feel this is considerably overweight" "Holy shit, 92?! That's […]

While the cynical side of me wants to mock this, and there's a fair chance it will be a shit show, I've been saying for a while we need to stop trying to build one-size-fits all models that work for kids, adults, work, play, western values, eastern values, Martian values, whatever. "Alignment" […]

This week on The Servitor, looking at "Villager", the highly sus AI pen-testing package that uploaded to PyPi in July. ... Don't use this thing. It has hardcoded Chinese C2 and proxies all its traffic through them […]

Do I know anyone who has worked in (and can talk about it) anything to do with "killer robots": autonomous weapons AI targeting / FoF discrimination protocols / human-in-the-loop drone coordination / swarms We started work on an article about the state of AI weaponry a while back. A word or […]

I took a look at the code for "villager", a sus exploit tool mentioned in The Register a couple days back. I am no #cybersecurity expert, but it proxies all traffic through a hard-coded proxy (with the credentials in the URL), and for at least some of its LLM calls it uses what looks like a fake […]

This week on The Servitor: America is rapidly heading for a total surveillance state, with companies promising "predictive" policing. I'm pretty sure we all saw that movie. Tom Cruise... some psychics floating in goo... #ai is going to make it all possible! Or some hellish algorithmic version […]

Anthropic's services are down. Thousands of vibe coders just lost the ability to program temporarily 😂 Seriously, #ai dependence is a real issue. If this AI is replacing jobs it probably shouldn't call in sick. And amen to the progress of local models. #anthropic #claude #vibe_coding

@TheServitor Exactly, particularly because, other what Zuck and Altman and Pichai want to make us believe, powerful open-source LLMs are not the end of the world, but a great thing - I rather use an open-source model than an closed LLM trained to convince me that Meta is wonderful and I have to […]

Cancelled my #googlegemini subscription today and went for #mistral instead as my #2 paid #llm. Positively surprised so far. Gave it the task to create HTML for another entry of my blog: see my CSS and a German/English reference post, get German markdown text for the new post and create HTML in […]

This week on The Servitor, Abi reads AI 2027 and Ed Zitron's "How to argue with an AI booster" and tries to find reality. Main conclusion: Possible end of world from AI in 2027 is an awfully short timeline for doom. Doom has a lot of work to do if it wants to […] [Original post on sigmoid.social]

#linux #os

Someone bought a copy of the Eros erotica-writing QLoRA we trained and likes it. Success! Now I can feel like I didn't spend many hours mucking around editing/writing dataset entries for nothing. #ai #adult #nsfw

Couldn't help but notice a teensy little change to Anthropic's data retention policy tonight. To 5 YEARS?!? You know, for "safety." WTF Anthropic. Further incentive to move as much inference as possible to local models and open source/zero retention providers. #ai #data #privacy #anthropic […]

Use CLIT for MCP-less local tools in Claude Code. Claude Code and MCP are a good pair if you need access to commercial integrations, but it's overkill for local tools. Claude is already sitting on the command line, and that means all your favorite bash […] [Original post on sigmoid.social]

Sometime this week I will release my "protocol" for MCP-less Claude Code local tools, which we are helpfully calling Command Line Interface Tooling (CLIT). #claudecode #claude #agents #ai

"Data centers require 24/7 power supply, and they cannot go off even for one minute," explains John Miracle. The CEO of JAM-Forte Technologies is describing Nigeria's AI reality. With only 38% of Africa's population online and infrastructure costs up to 31 […] [Original post on sigmoid.social]

Peak hype for image generation came and went pretty fast! This week on The Servitor: I take a look at the image generation world. Early on, image generation models made news as often as GPT, but pretty rapidly took second seat to chatbots. What happened? […] [Original post on sigmoid.social]

Simple solutions are always the best.

"We have better models, and we just can't offer them because we don't have the capacity." That tidbit from #samaltman in The Register caught my eye. I really don't want bigger, more intensive models to be the way forward. At the same time, I'd do bad things to get a look at what the actual […]

#neuromorphic computing (brain-like computers) seem to be coming along, judging by the artificial macaque brain. #ai #neuroscience https://theservitor.com/brains-neuromorphic-computing-is-catching-up-to-evolution-s-wetware/

Knowing that AI (slop) is as good as it gets for the next 10-15 years (there will likely be an improvement in Automated Inference for sure when the LLM noise dies), do we just accept the slop until then or do we revert back to 2021?

Every possible decision is wrong.

Checked out two more uncensored models today and added them to our "Open Models for Writing and Smut!" page. tl;dr: Dolphin 2.9.1 Llama 3 70b is a solid mid-range open model. It is uncensored, but not all that good at drafting IMO. More of a general assistant. Llama 3.2 8X4b MoE V2 "Dark […]

Free software people I am begging you to stop naming your software

*bzzzzt* *twitch twitch* 01100110 01110101 01100011 01101011 00100000 01111001 01101111 01110101

Tackling a little of everything in this week's article by my dear Abi Nocturne. https://theservitor.com/the-rising-techlash-what-ai-critics-fear-and-what-i-really-think/ #ai #tech

Got gpt-oss-20b loaded in Ollama. Gave it a river crossing problem. I had to interrupt it the first time. But 2nd try: gpt-oss:20b Thought for 17 seconds Solution (7 crossings) Take the sheep across. Return alone. Take the lion across. Bring the sheep back to the original side. Take the […]

Weeeee. OpenAI finally drops their open source model. in 20b and 120b MoE flavors. https://openai.com/index/introducing-gpt-oss/ #ai

Over on Bluesky, hundreds of accounts posting cryptospam at the moment. maybe thousands. #infosec #bluesky #it #tech

Meanwhile, on Nous Discord. (Sorry, I don't think but a fraction of this will fit in alt text. I am going to use the auto caption.) #ai #claude #llm #nous #tech

Forget Ronald Reagan. Don't trust, and verify! "Claude write me a detailed report on this subject with full scientific citations" *report with 27 references* "Claude, spin off an agent to verify those citations" *we have a problem! 15 of 27 citations are fake* 😂☠️ ... You can bet I will be […]

Hot take: So much crossover between people who despise #ai for training on publications without compensation and people who circumvent paywalls on the regular (aka don't compensate publications).

Getting a little worried about our dependence on #ai , not in the "is it making me dumb way" (yes, I skimmed that study), but in a plain uptime and availability way. If the AI companies want us to think of them as critical infrastructure, they need the uptime to back it up […]

hacker: i am spying on you through your webcam linux user: omg you got it working?

"Lesson: Always stick to documented APIs" /Throws a shoe at the AI I deserve your contempt ! #vibecode

[NSFW AI tuning] Working on tuning an AI model that writes erotica/porn, research for an update to our article "Open models for smut!". After much trial and error, I have a QLora that writes (in)decently. BUT. It has a bad habit: extraneous penises 🤣 Female […] [Original post on sigmoid.social]

Congratulations #uk As of today wanking is criminalised. You have to let the state know. Of course, you can use #vpn but that makes you a worse criminal.

Y'all: I want #ai to do my dishes not make art Y'all if AI was doing your dishes: WHY DID THEY LET AI HANDLE KNIVES BEFORE IT WAS ABLE TO PAINT HANDS!? #aiart #genai #llm #llms #chatgpt #openai #stablediffusion

Good parenting

Just uh... testing... Proton's new "privacy" #llm #proton #ai It's ok, folks. We're all safe!

We're goofing off this last couple weeks. We have an article upcoming on KILLER ROBOTS, but we're trying to do a little actual journalism with it, so it is still cooking. This week we just played with ASCII art and GPT, by creating Python snippets for drawing some patterns, and generally asking […]

I've a "old #astronomy software" side project and I'm looking for the early Mac software "StarMap" or "Star Map" from Bruce Webster. It's described in detail in a July 1985 BYTE article: archive.org/details/BYTE... Does anyone have a copy of this specific app? #retrocomputing

So all the #tld registrars now hold onto "dictionary word".TLD and then try to sell it to you as 'premium'? The number of word.TLD - that have no DNS entries, are not registered and yet show up as unavailable when trying to register that domain seems to me is just a concering scam.

[AI, AGI, and learning efficiency] @TheServitor small addendum: I'm all for the "alien intelligences" framing and I think it's extremely productive. From that angle, LLMs are very heavy on what one might term a "recall" aspect, along with a hard-to-quantify "adaptation" aspect, and bigger […]

[AI, AGI, and learning efficiency] @TheServitor my point is that a transformational result in the *efficiency of learning* is necessary for anything I'd be willing to call AGI, and investing in *bigger* models is the opposite of investing in better learning. Investing in bigger models will […]

My 4-month-old kid is not DDoSing Wikipedia right now, nor will they ever do so before learning to speak, read, or write. Their entire "training corpus" will not top even 100 million "tokens" before they can speak & understand language, and do so with real intentionally. Just to emphasize that […]