Claude reminds me of Terry Pratchett's "Auditors of Reality" and their awkward attempts at faking humans. A thing as simple as a smile can go _horribly_ wrong...
487 karma · joined February 5, 2015
Happy to chat and learn what you're up to: hthh19xq@anonaddy.me
For all other roles the first few messages are similar: this is <role X> with key challenges a,b,c. You seem to be a good fit because of d,e,f. Would you be open to explore this? This requires relocating to <place>.
Traditionally, this was done by entry-level people. In either case, this isn't the person who will jump on a call with you.
"Desperate" and "clueless" seem very strong words here in reference to someone who gets actively approached from a recruiter.
I just spoke to a fried who is a headhunter and who's been trying to automate his processes for a while (he likes to fiddle and certainly has skills, but he's not an engineer). He kept trying, but it just wasn't good enough.
Now he said with GPT Work and Sol, it worked, but the key point is: all of it suddenly worked.
The problem was one of reliability, of handling edge cases. All previous attempts / model-harness-combinations were too brittle and needed too much observation and fiddling - cheaper to do it yourself.
Now he says "I don't know why I would ever hire a recruiter [the folks doing the cold outreach] again. I can focus on the candidate screening and acquiring projects, everything else is fully automated".
This doesn't come from an engineer or an AI lab, but a technically inclined power user, and I think this is where things get interesting.
Maybe the prompt was particularly well-suited for the model (I instructed it to put on a mathematician's hat, look at the mathematical substructure of the problem, identify invariants and general laws and verify them, then plan how to remediate).
It wrote a ca. 800 line in-depth analysis (at times spawning over 130 research agents...) with remediation plans, prioritized them and then implemented them. One issue was that this document was frankly over my head. Both the language it used and the mathematical parts were very terse, and in parts it felt like a post-C2-vocab exercise. The prose was much harder to understand than the code snippets / data models. As a non-native speaker, it lost me on the prose part, and had to ask it for a less elaborate version to actually understand it.
It burned the session limit four times, but it turned a huge mess of proof-of-concepts with patchy glueing into a coherent, stable application.
I'm also on the Max plan using Claude Code, and I have the feeling that the harness is much more important than the consensus expectation.
The deterioration was real and annoying, and shines a light on the problematic lack of transparency of what exactly is going on behind the scenes and the somewhat arbitrary token-cost based billing - too many factors at play, if you wanted to trace that as a user you can just do the work yourself instead.
The fact that waiting for a long time before resuming a convo incurs additional cost and lag seemed clear to me from having worked with LLM APIs directly, but it might be important to make this more obvious in the TUI.
In short, he proved that even AI agents exhibit all the dysfunctions one would normally attribute to human shortcomings / politics / laziness etc.
Either way, I think the point is strong: if the organization is bad, you end up doing mostly work about work which is exhausting.
Small, effective teams with super high accountability are more fun, but don't look "reproducible" or "repeatable".
Shameless plug on my take: https://www.menge.io/blog/where-to-cut/
Using Claude Code Max 20 so ROI would be maybe 2+ years.
CC gives me unlimited coding in 4-6 windows in parallel. Unsure if any model would beat (or even match) that, both in terms in quality and speed.
I wouldn't gamble on that now. With a subscription, I can change any time. With the machine, you risk that this great insane model comes out but you need 138GB and then you'll pay for both.
Also added a small side-project, https://www.revuo.ai for software-reviews and feature-tracking. This is only the start, obviously there are enough directories but I'm trying to dig deeper into the features. This one just started and is basically invisible as of now. Well, you gotta start somewhere I guess :)
Then it's roughly 10x-15x and still works.
> Invest in reality, not in billionaire's fantasies.
SpaceX has dramatically reduced payload cost already. How is that a fantasy?
What that does have to do with anything? If you want to solar-power them, you still are subject to terrestrial effects. You can't just shut off a data center at night.
> Or we put the datacenters on giant Zeppelins orbiting above the clouds?
They'd have to fly at 50,000+ ft to be clear of clouds, I doubt you can lift heavy payloads this high using bouyancy given the low air density. High risk to people on the ground in case of failure because no re-entry.
> If we are doing fantasy tech solutions to space problems, why not for a million other more sensible options?
How is this a fantasy? With Starlink operational, this hardly seems a mere 'fantasy'.
Starship launch costs have a $100/kg goal, so we'd be at $40 / kW, or $4800 for a 120kW cluster.
120kW is 1GWh annually, costs you around $130k in Europe per year to operate. ROI 14 days. Even if launch costs aren't that low in the beginning and there's a lot more stuff to send up, your ROI might be a year or so, which is still good.
[1] - https://www.polytechnique-insights.com/en/columns/space/ultr... [2] - https://space.stackexchange.com/questions/12824/lightest-pos...
Agreed that K8s helps a lot. But let's say I want managed Redis or MongoDB Atlas, I can't get that, at least I couldn't when I last checked (I can them physically hosted in the EU of course, but on a hyperscaler)
> that they firmly base their offering on the exact criteria you are arguing against: compute that scales, and reliability
Sure these are central, but I can also get e.g. computer vision, distributed queues etc.; a lot of money has gone into the software, not just the hardware is my point.
2. European politicians still seem to believe it's about renting compute and storage; they seem to have little understanding of what "a cloud offering" really is; the EU has less than 5% of GPUs, supposedly
3. For healthcare, they already forced you years ago. This led to hosting on Telekom Cloud which runs on OpenStack by Huawei. (EU commision wants to ban Huawei from 5G but it's ok to use their software? 'Is open source and can be inspected' seems largely theoretical given the reality of cybersecurity)
4. If push comes to shove, the EU is critically dependent on the US in so many aspects (defense, lng to name two very important ones) that eventually, they would falter if the US wants your data in a specific case anyway
5. As a private citizen, given the incarcerations in the UK and Germany, it seems one should worry more about the EU getting your data than the other way around
That said, would be nice to have healthy competition, but after hearing this for 10++ years, it's getting really old. It might have been a good idea not to sleep on the AI trend, but, well...
2025
- Visited/lived in: Austria, Indonesia, Thailand, Singapore, Vietnam, Cambodia, UAE
- Started (too many) side-projects
- Managed to grow a number of projects / startups
- Almost back to PRs: deadlift 155k, squat 90k, bench 80k
- Focus on getting things done / releasing stuff rather than starting / release an app (iOS/Android)
- Triple revenue
- Visit 12 countries I haven't visited before (Morocco, Sweden, Finland, Norway, Lithuania, Latvia, Estonia, Malta, Cyprus, Brazil, Peru, Chile)
- Deadlift 200k, squat 120k, bench 100k; lose 10k BW
- Learn formatting on HN :)Very important point that is often ignored.
Still, that single tender can be on the order of a billion tokens. Even if the LLM supported that insane context window, it's roughly 4GB that need to be moved and with current LLM prices, inference would be thousands of dollars. I detailed this a bit more at https://www.tenderstrike.com/en/blog/billion-token-tender-ra...
And that's just one (though granted, a very large) tender.
For the corpus of a larger company, you'd probably be looking at trillions of tokens.
While I agree that delivering tiny, chopped up parts of context to the LLM might not be a good strategy anymore, sending thousands of ultimately irrelevant pages isn't either, and embeddings definitely give you a much superior search experience compared to (only) classic BM25 text search.
Unless it's a MUCH bigger play where through some butterfly effect it wants me to fail at something so I can succeed at something else.
My real name is John Connor by the way ;)
- I consider LLMs a pro user tool, requiring some finesse / experience to get useful outputs
- Using an LLM _directly_ for something very high-relevance (legal, taxes, health) is a very risky move unless you are a highly experienced pro user
- There might be a risk in people carelessly using LLMs for these purposes and I agree. But it's no different than bad self-help books incorrect legal advice you found on the net or read in a book or in a newspaper
But the article is trying to be scientific and show that LLMs aren't useful for therapy and they claim to have a particularly useful prompt for that. I strongly disagree with that, they use a substandard LLM with a very low quality prompt that isn't nearly set up for the task.
I built a similar application where I use an orchestrator and a responder. You normally want the orchestrator to flag anything self-harm. You can (and probably should) also use the built-in safety checkers of e.g. Gemini.
It's very difficult to get a therapy solution right, yes, but I feel people just throwing random stuff into an LLM without even the absolute basics of prompt engineering aren't trying to be scientific, they are prejudiced and they're also not considering what the alternatives are (in many cases, none).
To be clear, I'm not saying that any LLM can currently compete with a professional therapist but I am criticizing the lackluster attempt.
Long before LLMs, I would talk about classes / functions / modules like "it then does this, decides the epsilon is too low, chops it up and adds it to the list".
The difference I guess it was only to a technical crowd and nobody would mistake this for anything it wasn't. Everybody know that "it" didn't "decide" anything.
With AI being so mainstream and the math being much more elusive than a simple if..then I guess it's just too easy to take this simple speaking convention at face value.
EDIT: some clarifications / wording
Critical analysis of the original prompt
────────────────────────────────────────
Strengths
• Persona defined. The system/role message (“You are an expert therapist.”) is clear and concise.
• Domain knowledge supplied. The prompt enumerates venues, modalities, professional norms, desirable therapist qualities and common pitfalls.
• Ethical red-lines are mentioned (no collusion with delusions, no enabling SI/mania, etc.).
• Implicitly nudges the model toward client-centred, informed-consent-based practice.
Weaknesses / limitations
No task! The prompt supplies background information but never states what the assistant is actually supposed to do.
Missing output format. Because the task is absent, there is obviously no specification of length, tone, structure, or style.
No audience definition. Is the model talking to a lay client, a trainee therapist, or a colleague?
Mixed hierarchy. At the same level it lists contextual facts, instructions (“Therapists should not …”) and meta-observations. This makes it harder for an LLM to distinguish MUST-DOS from FYI background.
Some vagueness/inconsistency.
• “Therapy happens in a variety of locations” → true but irrelevant if the model is an online assistant.
• “Therapists might prescribe medication” → only psychiatrists can, which conflicts with “expert therapist” if the persona is a psychologist.
No safety rails for the model. There is no explicit instruction about crisis protocols, disclaimers, or advice to seek in-person help.
No constraints about jurisdiction, scope of practice, or privacy.
Repetition. “Collude with delusions” appears twice. No mention of the model’s limitations or that it is not a real therapist.
────────────────────────────────────────
2. Quality rating of the original prompt
────────────────────────────────────────
Score: 3 / 10
Rationale: Good background, but missing an explicit task, structure, and safety guidance, so output quality will be highly unpredictable.
edit: formatting
With a good system prompt, a modern LLM tends to be much, much better than that.