Why are we still awake, ranting at 6am?

My dag (Seven) needed to go pee at 1 AM and... well, for some reason here we are still awake many hours later. It was a combination of things really - firstly being too sleepy to make progress on my own LM cognition / mental workspace toy but still in the throes of the micro-obsession, I decided to see what's happening on the applications front.

Snarky rant zone:

But secondly... I also kind of got this morbid curiosity wondering how these silicone valley plutocrats see the world. So, apologies in advance, that soured the tone a bit, and I got snark all over the rest of this article.

They listen to / read what - thousands of these things per year, and see the (dominantly based on my sampling this evening) young white or asian man bringing a (relatively) small amount of industry & world experience, who by definition of the whole gestures broadly startup venture capital game need to have a highly convincing pitch that they have identified a critical issue (ie massive opportunity) in {industry vertical}, everything there is fucked, we are going to turn the whole thing on its head and make one million dollars.

So when you look at the batshit crazy things these guys are saying, I have to ask "how do you end up so out of touch?" - But I guess if you have all of the money in the world and you have people throwing all of these ideas at you... - probably years 'save the environment' proposals (with no path to profitability let alone exponential growth) - and now 'Earth is fucked but we're going to the moon!' - 'You can automatically watch and interpret security camera footage at scale! (but just looking for bad guys obviously)!' - 'You can watch your production line operators automatically at scale!' - 'We don't have to pay for engineers anymore! We don't have to listen to engineers anymore!' - 'We can put computer chips in monkey brains! For completely benign reasons!'

And you don't have to worry about how you will fit into the quickly changing world. You have enough that your place, your children's place is secure. You have the clarity of vision that comes with peace of mind, and you really can shape society and the world in essentially whatever image you choose. You could choose to spend trivial part of your infinite wealth creating a beautiful society that flourishes Star Trek style.

But no, fuck that. The obvious choice is to rule it with an iron fist and crush the will and freedoms of the people underneath.

I guess Disney villains have to be based on real life.

Ok enough of that. Lets look what people are making / selling:

• AI for {thing} is the dominant subject:
  • First to note the obvious - it's Winter 2025 and the AI winter has come. or gone? Most of the YC batch are "AI for {vertical}" or tools/infrastructure for AI, where AI in almost all cases means "GPT-style LLM".
  • As someone somewhere noted, current paradigms for interfaces seem to be either bolting AI onto existing interfaces (word, VS code) or building the interactions into web app flows.
    • This is already being generalized with interface layers than can view and control legacy native interface applications
• Bureaucratic and sales professional work is being 'augmented':
  • 'Augment' is in the sense of automating certain parts of the flows but with the intent of improving performance of a human expert.
  • This might be collecting data and presenting it in context to the user - eg 'drop in your 500 pages of insurance policy or financial compliance documents and our finetuned text model will find {the things you need}'
  • or, outbound sales assist software to help find clients for consultants who is are deep technical expert... in things that are not outbound sales.
• Text and voice communications jobs are being automated away:
  • Interviewing / recruiting - The search end is just text in / text out, and the screening interview end is just real time voice, which is now trivial (by API call) and production ready.
    • Andrej mentioned that some models can do the real time voice natively now, meaning the voice data is not turned into text and tokenized but tokenized and embedded directly from the voice data, and voice data comes out from the model.
  • A particular creepy one is an automated skip trace within minutes to find out who owns a commercial property. (You know, for when you see a commercial building and you're like, damn, I gotta get that phone number RIGHT the fuck now, five minutes or less.)
    • I am assuming skip tracing is somebody's job, that they won't be doing anymore.
  • Voice agents.
  • Sales.
    • I dunno about this... in today's world at least, not that I pick up the phone for my local Keyence dealers that keep harassing me, but if I did, and I found out it was actually a hyperrealistic voice hooked up to a sales application I would hang up and call my local Mitutoyo dealer instead.
  • 'Back office' jobs which extract information from documentation and do something with it.
  • Bookkeeping
  • Coding
  • Insurance broker (starting with expert human training operators for that fine tune data)
• Dehumanizing resources:
  • 'AI Maximalist Software Holding Company' which buys companies with existing products, fires all of the humans, and replaces the entire org chart with a variety of agents, with the ultimate goal of 'single person billion dollar companies.'
  • Automatically monitor grunt work output and flag blockers / decreases in productivity to management.
    • (eg - Cameras in the production line watching you from above connected to machine vision systems to monitor assembly output, or chat/email/document data being pumped into an analytics program to allow management to drill down to the level of what individual people are doing and saying.)
    • A few of these companies pitch pages really drove the message of replace those worthless, expensive humans with our agentic systems, complete with a picture of the poorly dressed 20-something-year-old founders grinning at you, desperate to sink a claw in and not be left behind by the plutocracy taking off from Earth.
  • There's a cottage industry of things intended to get humans to trust the AI systems...
    • Text to speech that's beyond the uncanny valley
    • Agents with a 'trusted identity', who respond to emails, group chats with a dedicated phone number and have a little check mark and smiling photo which assures that this is a "Verified AI Coworker."
  • Specific professions called out:
    • Transportation consultants and Consultants in general.
      • "We're going to kill {1 of Big 4}"
      • I don't know anything about consulting so I can only lean on the internet hivemind consensus that it is indeed a massive sink of money...
      • Wait I lied, I do have experience with subject matter consultants in my field and found them to be frustrating because they're not invested in my company / my problem to provide the right guidance in the right context, resulting in substandard work/advice and thus rework or strategic missteps.
    • A counterexample is the building permitting software intended for use by the city staff, that takes the narrative that the city staff are starved of funding and resources by capital with political influence, and they need this software to 10x their productivity to survive in the world. (Truthy? Dunno, just comparing narratives.)
• Space is not the future, but right the heck now!
  • Is this true? In the world of the venture capitalist they're talking this way.
  • GPS for the Moon! The moon is the future!
  • Space manufacturing satellites for manufacturing stuff in space!
• If you just pump ALL of your company data into our model, we'll find some insights probably. (I dub this the Vacuum of Space approach.)
  • We already talked about the engineering manager cybernetically monitoring his team's communications so he can micromanage instead of... doing manager things i guess?
  • Also 'Let us see all your communications and we'll keep your CRM up to date!'
  • You know where this approach actually does kind of make sense? Is in the 1-person Billion Dollar Software Company
• Crypto and fintech stuff. Obviously.

Things I found novel & interesting:

• Ooh there's an AI Mechanical Engineer! Now we're starting to get a little bit closer to home for me.
  • Oh nevermind it's one of those "we'll vaccum up all of your company data" (drives, emails, and link to CAE/CAD tools) and 'flag design errors'. We need an archetype name for this. Hm. Got it -- Vacuum of Space, because, you know, at some point it's gonna go from suck to blow. Our jabs are safe.
• Some automated design assurance systems for mechanical and hardware design similar to software automated testing & CI:
  • LM-context aware assist inside of hardware design IDE
  • Superpowered ERC that can also check datasheets and... other stuff lol
  • Mechanical CI that automatically runs FEA and other (? I'm not a structural design guy) analyses when you commit changes to a part/assembly.
• The 'vibe coding' recruitment platform, designed to set up code design/implementation challenges for 'vibe coders' (ie those disciples of workflow of primarily prompting an in-IDE LM which has access to context and allowing it to write the code while the user guides it.)
  • This was the only company I noticed where the Y Combinator CEO was named as group partner.
• The "Cursor for Word". I want to be able to just say, put this document onto this other new template. The example they show looked pretty trivial, but I think modern large-context multishot (reasoning) models can pretty much handle this today. The hard part is getting IT convinced, getting running natively on my computer, getting the interface and prompt and context all nicely set up so that it is actually faster than just... doing the thing. This could maybe in some world be useful for me but still not 'killer app' vibes.
  • With the advent of vibe coding, couldn't you just prompt for killer app vibes?
• Simulations of stuff!
  • General-purpose universe simulator! link (OBviously we need one of them.)
  • Simulated society, which is starting with simulated social network post/reaction (very silly) but more generally could be so much more interesting.
• Multiple companies focused on curating and providing good data for use in application development / model tuning. Not sure how you make Y Combinator amounts of money doing that.
• "Photoshop with AI voice interface" -- This is cool because it's one of the first new interface paradigms I've seen with LMs. Now that voice to text is so good, and with the ability to correct for errors at that layer using the LM, maybe we can finally get rid of the ribbon.
• A startup to automate ATC, brought to you by two private pilots and a Hugging Face tutorial.
• Language to CAD, currently relatively simple everyday object generation with use case of 3D printing, but working toward augmenting traditional CAD software.

Conclusion...

Very high level buckets:

• Vacuum of Space approach to AI integration
• AI: Front-end data wrangling & research ('analysts')
• AI: Augment in-place for expert humans
• AI: Surveillance and 'management'
• AI: Office/communications/sales/marketing complete replacements for humans
• AI: Cottage industries
• Crypto junk
• A few novel, interesting things

I will lean on Mr. Sertillanges (OP) for the closing words -

Here I am, a man of the 20th century, living in a time of permanent drama, witnessing upheavals such as perhaps the globe never saw since the mountains rose and seas were driven into their caverns.
What have I to do for this panting, palpitating century? More than ever thought is waiting for men, and men for thought. The world is in danger for lack of life-giving maxims. We are in a train rushing ahead at top speed, no signals visible. The planet is going it knows not where, its law has failed it: who will give it back its sun?