NEW! ScottiesTech.Info presents: TekTalk

[Scottie]

Behold, episode 1 of:

TekTalk with Scottie & Toño

This idea has been simmering for quite some time... It came to a boil this year after several things Toño posted or sent to me made me go, "Oh, that would make a GREAT video!"

Still working out the kinks with audio/video and trying to find our "optimal flow"... Even though I've published 270 videos, I was quite nervous. Silly, but true.

Let us know what you think.

 

[Claus]

Thank you @Scottie, it is a good format, I like it.
Maybe you could put the logo on one of the top corners, I think that would look cleaner.
I'm not ful through the show now, but I can't wait for the next one. :D

Edit: Maybe you could also look, that you are not mirrored.

 

[Scottie]

RSS feed for podcast format:

https://media.rss.com/tektalk-with-scottie-tono/feed.xml

The podcast should also appear soon on Apple Podcasts, Spotify, Amazon Music, etc.

 

[luc]

RSS feed for podcast format:

https://media.rss.com/tektalk-with-scottie-tono/feed.xml

The podcast should also appear soon on Apple Podcasts, Spotify, Amazon Music, etc.

Nice, for that kind of format I prefer listening in podcast form while doing other things (and that's probably true for many people).

Two minor suggestions maybe for the future: some kind of USB mic (as opposed to lavalier) would probably improve sound quality, which is especially important when listening on a crappy phone speaker like I do. (Not that it's bad at the moment.)

Second, I noticed with podcasts that people are using very click-baity titles these days, and to be honest, they seem to work (at least on me!), even when the podcast is actually softspoken and nuanced. Seems people are just used to it at this point. If you want to go down that route, maybe bold statements instead of questions would work better, like "Agentic AI Will Change Everything (And Make It Stupid)" or whatever bombastic claim you can cook up

 

[Scottie]

I'm using the same Rode SmartLav mic as for my other vids - plugged into a USB audio adapter though, not the usual Zoom digital recorder.

Toño's audio is quite echo-y...

 

[Scottie]

Apple Podcast:

TekTalk with Scottie & Toño

Technology Podcast · ScottiesTech.Info presents: Scottie and Toño talk tech! News and views on modern tech that you won't find anywhere else. Find us at: ScottiesTech.InfoSubstackYouTube

podcasts.apple.com

 

[irjO]

Awesome! As a “tech geek” myself, I like these kind of videos/podcasts! Ep 1 is great! Keep it up both of you!

 

[KJS]

Enjoyed the talk! But wow, you guys are conservative I find it really enjoyable to run models on Apple Silicon locally and extend them via various MCP services like web search, memory, etc. With smaller models, limitations are very pronounced, but they are still powerful enough to extract data from written text, perform NER, or other NLP‑to‑structured‑data processing.

As for coding, I haven’t written anything by hand for two or three months. It’s not possible to have everything working in one shot, and providers often route to heavier quantized (=dumber) models during peak hours, but with an automated review process it’s manageable. For instance, I was able to port some legacy Postgres/Rails reporting code with stored procedures to a clean ClickHouse big‑data queries (on S3‑stored data) architecture. I wouldn't be able to tackle this without psychiatric meds and months of plumbing... So in many cases, those tools make software development actually pleasant.

Are you planning to stick with the current one‑on‑one format, or will you be inviting guests in the future? I'd be really interested in seeing your cold‑headedness juxtaposed with someone more all‑in on this stuff (like Andrej Karpathy?).

 

[Yas]

Good job you guys! It was an interesting talk.
I'm with you in the scepticism about AI. I have used it for several things and I usually don't find it very impressive. It can save up some work in small tasks but I feel it usually needs A LOT of checking and tweaking.

What I was thinking, though, while listening, is that perhaps most people don't really want something super impressive. I mean the general public of course, not the more nerdy ones.

What I mean by this is that people I talk to seem to be quite impressed with Chat GPT and Grok. And they use it for all sorts of things already in a way that I find incredible. And then you have all this new 'content creators' who specialize in teaching people how to prompt, and I have spoken to a lot of people who actually join these 'courses' about 'how to create prompts for chat get' I mean, it's amazing. So, that's my thought, that perhaps as it is right now it's got people quite impressed already, and perhaps the big companies are counting on that too.

Anyway, regarding the show itself. I think the echo isn't too bothersome and the flow is nice, so keep it up!

 

[Navigator]

What I was thinking, though, while listening, is that perhaps most people don't really want something super impressive. I mean the general public of course, not the more nerdy ones.

Nop, not the regular folks, usually convenience needs to be there first. The people using OpenClaw are these early-adopter power users types, and people who definitely need a helping hand LOL

 

[Scottie]

I confess I'm pretty anti-AI, but only because IMO it's simply not good enough. It is of course possible that my expectations are just too high, but...

I just used it again tonight for a simple issue (printing problem) and the solution it gave was to install an app on the Microsoft Store that doesn't exist. A simple scan of the manufacturer's web site shows there is no MS store app! I mean, c'mon!!

And as for "context files", well... imagine using a search engine where you have to tell it the results so that it can give you the results you want/need. That just doesn't make sense to me.

I demand nothing less than the ship's computer on Star Trek: The Next Generation!!

 

[KJS]

I just used it again tonight for a simple issue (printing problem) and the solution it gave was to install an app on the Microsoft Store that doesn't exist. A simple scan of the manufacturer's web site shows there is no MS store app! I mean, c'mon!!

May I ask which one you are using? For troubleshooting, Google Gemini with Gemini 3.1 Pro and its web search rarely gives me an incorrect answer. Even when I tried to find where I should click on the electronic tax office form, Gemini gave me correct answers. Apart from that, I'm not sure if tasks like these are the ones where LLMs shine (although they are marketed differently). There's a lot to the system prompt and how the model was tuned. I found out that smaller models like Haiku or Gemini Flash are tuned more for instruction following to compensate their dumbness. This means that if some RPC call (for ex. web_search) fails and there's no clear room in the directives to give a fallback answer, they try to use their limited knowledge to provide any answer.

I think that better uses are translations: language to language, business domain to code, code to code, multimodal data extraction (not only OCR but handwritten text), and text‑corpus labelling (not limited by short context sizes of transformers such as BERT). Those tasks can be done locally on Apple Silicon powered machine with 20W peak usage.

 

[Scottie]

May I ask which one you are using? For troubleshooting, Google Gemini with Gemini 3.1 Pro and its web search rarely gives me an incorrect answer. Even when I tried to find where I should click on the electronic tax office form, Gemini gave me correct answers.

Grok in this case. Also tried DeepSeek, Gemini, etc. None were paid versions, though.

Apart from that, I'm not sure if tasks like these are the ones where LLMs shine (although they are marketed differently).

That's one of my main issues: they are being marketed as Intelligent Universal Jesus Software, yet the majority of them really aren't that good.

There's a lot to the system prompt and how the model was tuned. I found out that smaller models like Haiku or Gemini Flash are tuned more for instruction following to compensate their dumbness. This means that if some RPC call (for ex. web_search) fails and there's no clear room in the directives to give a fallback answer, they try to use their limited knowledge to provide any answer.

The day it answers: "Scottie, I have no idea! Try the other guy, he's good at that stuff!" is the day I will praise AI.

I think that better uses are translations: language to language, business domain to code, code to code, multimodal data extraction (not only OCR but handwritten text), and text‑corpus labelling (not limited by short context sizes of transformers such as BERT). Those tasks can be done locally on Apple Silicon powered machine with 20W peak usage.

For language stuff, it's fantastic. Transcribing audio, video, generating SRT files in multiple languages... it's super quick, accurate, and easy. I'm still waiting for my Star Trek Universal Translator comm badge, though.

Haven't used it for coding, and for the most part I don't want to - yet.

I'm still waiting for the AI Hype bubble to pop, and then I think we'll see many new amazing things that do not require massive data centers, nuke plants worth of energy, and 300GB of RAM.

I did the same with "Web 2.0"... First there was tons of hype, then the bubble burst, and a few years later I really dived in. But then of course the control and censorship also kicked in/ramped up in a huge way, so... We'll see!

 

[KJS]

I'm still waiting for the AI Hype bubble to pop, and then I think we'll see many new amazing things that do not require massive data centers, nuke plants worth of energy, and 300GB of RAM.

I did the same with "Web 2.0"... First there was tons of hype, then the bubble burst, and a few years later I really dived in. But then of course the control and censorship also kicked in/ramped up in a huge way, so... We'll see!

I think this is already starting slowly: https://www.kickstarter.com/projects/tiinyai/tiiny-ai-pocket-lab. I guess this will result in small, for‑task‑tuned models, taught (distilled) from frontier ones (which, technically, is forbidden by the ToS).

The general problem in the AI coding space right now is inference instability and rate limiting. By instability, I mean that models are deliberately routed to heavily quantized versions during peak hours. There are even websites monitoring this, such as AI Benchmark Tool - Best AI Models 2026 | Compare Claude vs GPT vs Gemini. Quantized versions can easily lead to bad design during task planning and/or a crazy-like implementation, or a model that gets stuck in a Tourette-like loop, burning through tokens. Just yesterday, GPT 5.3‑Codex decided not to generate code diffs for me(which it does absolutely great), but to… generate a Python script that patches the source code for every change and executes it.

Rate limiting is even worse: even if you purchase the Claude Code Max x20 plan, marketed for heavy usage and costing $200, you often need two or three Max x20 plans, which can lead to Anthropic banning you. You can easily get rate‑limited after an hour or so of work, "just because". This is causing LLM self‑hosting to start booming, and people are coming up with new ways to run them without needing large providers and siphoning all the data to them.