threecheese

I’d think any company big enough or working in certain markets which has a Compliance Officer cares about this; regulations are a legitimate business risk, and software integration contracts have security control compliance requirements which very much impact the sdlc.

Would you have the same reaction to requiring an approval for a production deployment? That’s driving the development process.

—-

Also jfc I need to cool it with the buzzwords, sorry I just got home from “talk like this all day” $job

Because they are selling a “pro” version as part of their commercial product SlicerVM. It has more features for operating a k3s cluster.

It’s even worse when you cant push back with “did you even read this”, because the politics haven’t evolved to constrain the slop.

We are getting pre-solutioned massive epics, dozens of files, from senior leaders (non-ICs); when shit goes sideways, what do you do? Our jobs are already at risk just in general, and we have new KPIs around generative AI (as do those senior leaders). I’m not sticking my neck out get chopped off.

Just last week I had to make some shit up in my uplevel status report to shift blame away from an AVP. Technically it’s my fault, for not digging into the 30 files (and tanking my own metrics,m); I don’t even feel like it matters - the devs just hand that off to an LLM anyway to meet their KPIs. I’m just thankful it didn’t go to prod.

You are out of your mind, and I love it! base32768?

This appears to be Claude (et al) run amok on a problem, the solution is very creative and well detailed. I had my Claude take a crack at examining more character encodings, to see if it could be optimized for varying input URLs:

> The two-tier base32768/base91 dispatch is the right design because the output alphabet and the input character set are independent concerns — the encoding choice should be driven by the two metrics that actually matter (visible chars vs. wire bytes), and base32768 and base91 already represent the optimal endpoints of that tradeoff. Any intermediate alphabet would be dominated on both axes simultaneously. Meanwhile, input character distribution is already exploited implicitly by deflate’s Huffman stage, so there’s no compression signal left for an explicit tier system to capture. The architecture is both minimal and complete: two tiers, zero header bytes, and no complexity left on the table.

Agree. My point is that other secret should have been inaccessible without an escalation. The fact that it was available to the agent implies a lack of basic security controls; in fact I would expect that an agentic workload would have even more robust compensating controls.

If anyone is interested in a peek into why they are being so aggressive, check the “AI Hype” board [1]; beyond all the interesting local models (why I read it), it is usually filled with projects for circumventing LLM provider restrictions which are wildly popular (and frequently Chinese- no shade).

The #3 result today is: “End-to-end protocol replay toolkit for ChatGPT Plus/Team/Pro subscription with from-scratch hCaptcha solver and empirical anti-fraud research”. The “research” for anti-fraud is “how to get around it”.

It looks a lot like an arms race, and we are getting caught in the middle of it.

1. https://hype.replicate.dev/

My understanding is that Rivian’s lane keeping (and other features like it) are only possible because of driving data collected to train their models.

It’s not such a stretch to believe that there’s some aspect of this that is specific to a driver or to a vehicle, and so requires that they collect your data. Even if this is not accurate, I can see a business making the decision that, given they need more and more data to improve the model, they would not allow customers to opt-out of that training cohort and still use the feature. Incentives etc.

Directionally though, I am with you on auto telematics data collection; I am not sure you can even buy a new car in the US that doesn’t ship with tracking, and many manufacturers (like the one who makes my car) don’t allow opt out at all. Fcking Stellantis

They have interesting features, their initial release had snippets, team sharing, time-saver stuff with a nice UX. Same reason one might use Raycast. I was a paying customer for that release, but when they pivoted I cancelled.

There was an early product named Rewind.ai (they have pivoted to Limitless Pin) which did essentially the same, I used it frequently - it essentially browser history but for everything you do on your computer, and you’ll get just about the same value as that.

We’re not normies, so take that with a grain of salt. Here’s mine: apps have access to significantly expanded capabilities which has privacy implications. If I can use the browser for a given app, I do it. Amazon for example.

Per their docs they have both “account” tokens and role-based tokens; the former have wide latitude (and might be used for DNS or root-access type stuff), while the latter are intended to be used for maintenance and have strong security boundaries. OP gave access to the former type without realizing it.

In most orgs, those would be behind some escalation control. Unless the token creator didn’t know what they were doing/creating, which tracks for a non-expert.

And then you get friends like Claude - dozens of them - which prefer to crap all over $HOME.

The app ‘Conductor’ does this, and I had to uninstall it; I just can’t crack my ‘ls ~/.co<TAB>’ habit, and “nd” is juuuuuust ahead of “nf”.

It *used to* be ‘~/c<TAB>’ before .claude crapped itself into existence..

This is the most interesting idea imo; do you think it’s testable? For example: allow the installed app to persist, turn on notifications, do some stuff to let a queue drain. Then remove.

Thanks for the wonderful Wikipedia excursion I just enjoyed, I learned a lot.

Anecdotally, I get the same wall time with my Max x5 (100$) and my ChatGPT Teams (30$) subscriptions.

It reminds me so much of the Underpants Gnomes (from South Park):

1. Use AI 2. 3. Profit!

Almost like that’s his job.

Hey, I’m with you - I think social media needs to die specifically for this reason. I’m reminded of the term “snake oil” - it’s like the dawn of newspapers again.

Any reason why you’ve closed access? I could swear there was more here yesterday

I’m jaded, and I shouldn’t let it show :)

One only needs to encounter a handful of bad actors to paint everyone with that brush - perhaps unfairly. Best of luck to you.

It is generalizable given a defined ontology, even better if your life experience can be represented using my ontology.

I had a large context model analyze the last ten years of my notes and build an ontology, it took a lot of iterating. Examples; A software project may have Decision, Risk etc entities; Life may consist of Activities, Goals, Concerns, Problems, etc; the World has knowledge/facts (Topic Taxonomies like Wikipedia) etc. These are all joinable given the relationships are intact.

The agent put everything into a huge RDF ontology - a world model. I worked with the agent to re-frame that large ontology, to I can build a skill appropriate for a small context model to serve as an expert on the ontology, it owns it for all intents and purposes.

I then worked with an agent to define use cases using my notes, real world things I do and have done: research, project management, goal setting, “hey I found a cool project online, it would be useful for my X project which is on the back burner but I don’t want to forget about it” you get the idea. I used those use cases to build out a few skills.

These serve as the actual ontology-aligned data layer, and so have access to URIs pointing to Goal entries in Obsidian, Projects in ClickUp, my calendar etc etc. And so it knows what my most salient Concerns are, and which other entities are associated with them - projects, goals, documents etc, and utilize various MCP tools for external systems. It also creates “context packs” using its view across systems, and so I can have it export structured markdown for any arbitrary entities and their direct and transitive relationships, to facilitate some other agent performing targeted work.

This enables me to: 1. Build targeted software tools for life and work management, I bring the skills into that project which own the data layer OR export context in the form of agent-ready markdown 2. Give skills to my Claws (which aren’t Claws at all, it can be any arbitrary skill supporting agent harness like Goose or Hermes) and LLM apps like Claude Desktop.

Totally agree, but I would say that strategic thinking is easier for the wearer of the boot than the owner of the neck.

Said less calamitously (word?): while it’s important to be objective, objectivity is difficult when there are real existential risks.

Thank you for the frank discussion!

I dislike the “brainwashed” comment from sibling, I believe it makes some assumptions. There aren’t any doubts that:

- AI is extremely resource intensive, consuming electricity, water, silicon, etc at levels possibly never seen before in humanity’s history; whether that’s a waste or not is subjective - Massive datacenters are popping up like anthills, and coupled with R-flavored regulation rollback there is a definite risk for environmental impact - just like during our last industrialization push where we poisoned much of the country, leading to a massive rollout of environmental protections in the 1970s and 1980s - Students are taking advantage of LLMs to shirk school responsibilities. Whether this is damaging or not is subjective until proven, and AI may not be causal here (students may not have been getting the expected value from their education without LLMs, again remains to be proven) - Many companies have used AI as a justification for layoffs, who knows what’s actually true though. There is a very real fear across society that it will continue to impact jobs, and senior AI company leaders are fueling this with public predictions of massive labor shifts. Again, maybe they are lying, but can you blame anyone for worrying?

There are counterarguments to all of these, but dismissing the fear as uneducated or brainwashed reveals your own priors and ignores all of these facts. It’s healthy to ingest OP’s criticisms - especially on a form populated most by Smart People (tm).

Like you, I dislike that the providers make it intentionally difficult to retrieve conversation history from their web UIs; you can Copy/Paste easily, or use the OS Share feature to access a public version, but they make it very difficult for me to build tooling to extract the history - it requires website automation, or a browser plugin.

I have many more fears than just annoyances :)

My biggest annoyance is hiding Thinking tokens; I have little trust in these aliens, and seeing how the sausage was made helped me to be more comfortable with eating it. Anthropic was the biggest provider that did not do this until recently, and they give a good rationale for the switch but that doesn’t make it less annoying. I also dislike the UX they put around it, “Hmm.”, “I should think about this” etc.

I do this every day, because Codex writes my requirements and Claude implements them. Just ask it for whatever you think the next model will need, tell it to be verbose if you like, and even have a second ChatGPT check it if you are worried. You can even give it a format, going as far as providing a specification or template if you do it frequently. Stick that template in both ChatGPT and claude projects so one can write it and the other can read it.

Edit: I shouldn’t admit this, but I even have an ontology defined - RDF and all - for some of my LLM tasks. Its classes contain examples, and so is like a few-shot instruction, and it’s working scarily well for structuring tasks.

I’ve been getting a lot unsolicited “hey, I saw your github” AI-generated emails lately; I appreciate your post for revealing the rent-seeking “talent-industry” middlemen behind this. I’ll go anonymize my github profile to hopefully reduce this - and I quote - “Unsolicited Commercial Email”.

Folks, this is how the death of SaaS is going to impact us; commercial interests are going to find every crack in your personal privacy into which they can wedge a vibe-coded automation.

Not sure if I’m a user just yet, but this quote stands out for me:

> If a node isn't physically standing next to you, it doesn't exist on your network. As humans move throughout the physical world, the data propagates organically.

/applause

Start with booze; always works :)

OP’s Qwen3.6 27B Q6 seems to run north of 20GB on huggingface, and should function on an Apple Silicon with 32GB RAM. Smaller models work unreasonably well even on my M1/64GB MacBook.

I am getting 10tok/sec on a 27B of Qwen3.5 (thinking, Q4, 18GB) on an M4/32GB Mac Mini. It’s slow.

For a 9B (much smaller, non-thinking) I am getting 30tok/sec, which is fast enough for regular use if you need something from the training data (like how to use grep or Hemingways favorite cocktail).

I’m using LMStudio, which is very easy and free (beer).

Go even further, and add this into the skill-creator skill, and let the agent improve the skill regularly. I do this with determinism, and have my skills try to identify steps which can be scripted.