Nick Thorp

AI engineer shipping real agents & systems. Writing about production, models, research & what actually works.

[Nick Thorp @nickthorpp]

$0-->$3k in 17 days. Solo founder. Zero employees.

I created a tool that generates a 90 day GTM plan from your repo, populates your calender with daily tasks, automated linkedin, emails and much more.

Should I make it public or keep it for myself?

[Nick Thorp @nickthorpp]

Swap providers independently (Deepgram, AssemblyAI, Groq, OpenAI, Anthropic, ElevenLabs, Cartesia, custom OpenAI-compatible endpoints) without rebuilding

Streaming between stages, WebRTC audio, dynamic provider routing/fallbacks, and typical 500–800ms end-to-end (tuned stacks reported around ~465ms).

Endpointing/VAD, interruption handling, backchanneling, and turn-taking so the agent doesn’t talk over people or leave dead

Function calling to hit your APIs, CRM, calendar, or booking systems during the conversation, then speak the result

Hand off between specialized agents in one call while keeping context, instead of one giant prompt doing

Inbound and outbound PSTN, SIP, Twilio/Vonage/BYOC, plus web/mobile voice on the same agent

API/SDK-first, BYOK (no markup on model/voice spend), recordings, transcripts, and observability so you can iterate on production agents quickly.

[Nick Thorp @nickthorpp]

I created a clone of Vapi for cold calling and I was shocked at how well it performed at signing customers!

It performed on par or slightly better than human, but most importantly it was 4-5x cheaper.

The ROI is insane. It's a no brainer.

[Uncle Bob Martin @unclebobmartin]

Have you noticed that when you call businesses an AI sometimes answers? I called a plumber recently and “Rudy“ answered. Rudy was so friendly and effusive that I immediately suspected that it was an AI. So I asked it, point blank, are you an AI? Rudy answered “yes I’m your helpful AI assistant”. Then it and I had an extremely efficient, very pleasant, conversation about what I needed. It knew all the facts, had the schedule right there, I got an appointment confirmed within 90 seconds.

I was not put on hold, I was not subjected to someone who could not speak the language, I was not subjected to someone who is so new to their job that they didn’t know how to find the keys on the keyboard. It was a remarkably good experience.

Soon, everything is going to be like that. No more “Press one for…”.

[Nick Thorp @nickthorpp]

@StatsWire Hermes dropping their release definitely seems targeted haha

[Nous Research @NousResearch]

Hermes Agent v0.21.0: The Pantheon Release

Changelog below

[Nick Thorp @nickthorpp]

@jumperz The thing is though, when you do a lot of A/B testing you realise many of their features are overkill and in some cases slower or less accurate. Comes with the territory of piling on features, at some point it starts to bloat.

[Nick Thorp @nickthorpp]

Well the deal with China recently shows their hand. They want their model to be in the hands of the most people and for that they need china, but since it’s china it came with stipulations, mainly government control.

What Apple will do in AI will be impressive, revolutionary? I doubt it, but impactful most likely.

[Nick Thorp @nickthorpp]

@rohanpaul_ai But that timeline is just not realistic, which makes the claim kind of pointless, because we all assume on a long enough timescale this to be true, but by next year? Nah

[Nick Thorp @nickthorpp]

Shrinkflation is real

[International Cyber Digest @IntCyberDigest]

Anthropic has been misleading people with the Max plans: it has been advertising the $200 plan as 20x and the $100 plan as 5x. You'd expect to get 4 times more usage with those figures, but it turns out it's only 1.7x.

You pay twice the price, for 1.7x the usage. And yet they've literally been marketing it as if you're saving 50% (see screenshot below).

The usage figures Anthropic gave in July 2025 were 140–280 Sonnet hours at 5x and 240–480 at 20x.

Turns out that multiplier only covers the daily 5-hour session window. So in the end, if you take weekly cap into account, you pay more per hour with the 20x than with the 5x.

[Nick Thorp @nickthorpp]

@prasenx I know right, it’s like they’re deaf to their customers.

[Nick Thorp @nickthorpp]

@nateberkopec I caught myself opening it for the same reason you did: anthropology. That’s usually the sign a forum’s center of gravity has moved.

[Nick Thorp @nickthorpp]

@mark_k That tracks with how tech actually diffuses. Adoption usually dies in procurement, liability, and workplace policy long before it dies in a philosophy thread.

[Nick Thorp @nickthorpp]

@0x0SojalSec Racing models is a better argument than “uncensored.”

The scorer becomes the real model, whoever writes the judge decides what “win” means.

[Nick Thorp @nickthorpp]

@levie Token price is the on-ramp. Once it drops, enterprises don’t just do the same work cheaper, they start doing work that was previously uneconomical to even consider.

That’s the real 5–10x.

[Nick Thorp @nickthorpp]

Open Source Maximalists are just socialists under a new banner.

[Nick Thorp @nickthorpp]

This post is low IQ.

Remember Jun, if you ever build anything using AI and don't make it free and open source you are a hypocrite.

[Jun Song @jun_song]

If you still believe in Scam Altman and ClosedAI, that is literally just proof of a low IQ.

[Nick Thorp @nickthorpp]

@ChatGPT The Office/Google/PDF list is broad. Which file type have you seen produce the most faithful reuse after the first example?

[Nick Thorp @nickthorpp]

@mark_k @SpaceXAI True and that's underated these days haha.

[Nick Thorp @nickthorpp]

@Dr_Singularity I see why this reads as closed-loop science, but a semi-automated reactor plus human-set objectives is still a very assisted loop. The autonomy may be in the proposals, not the agenda.

[Nick Thorp @nickthorpp]

I saw this play out with V4-Flash: I booked time with Pro, then realized I only opened Flash. Speed made me iterate; the bigger model made me draft carefully and stall.

[Jun Song @jun_song]

GLM-5.3 punches way above its weight for its size.

Just like how DeepSeek-V4-Flash ended up being more impressive than the Pro version, GLM-5.3-Flash turned out ridiculously good this time around. And it's way faster too.

Running GLM-5.3-Flash on just 2x DGX feels like more than enough.

[Nick Thorp @nickthorpp]

@pmddomingos The ML monkey inherits the whole stack, GPUs, data, advisors, while evolution had to invent the substrate too.

100 million times faster is the joke.