Felix Thöni

AI trainer @WIFI_Tirol , AI optimist, CTO at CTO at Building THE AI OS

[Felix Thöni @f_thoeniX]

@riuros @kimmonismus *mythical ways*

[Felix Thöni @f_thoeniX]

@kimmonismus It's time we slowly replace all EU politicians with AI. I advocate for GPT-3.5 - it is slightly more coherent than any of our politicians and we won't notice a difference.

[Felix Thöni @f_thoeniX]

GLM is ahead in terms of physics, for pure design (3d Tasks) GPT should be better considering the training data OpenAI uses.

I have been coding a majority of my full stack app [that is in production, and has a multi million euro ARR, so the stakes are high] on my own with Claude in multiple systems (claude code, Cursor, VS Code...) I have used Opus ever since September last year as my go to model for programming. Last week since it came out I started GLM 5.2 and I have to say: I built more features in a week than I did in a month with Opus 4.7 ... The crazy thing is that neither me nor my team had any issues with features we made with GLM which is a massive step up from Opus (but we were even more impressed and faster with Fable).
For now while Fable is not available, we'll be using GLM and even with Fable being available again we've found some uses for GLM that it excells at and beats gpt and claude / grok by miles in: the fine tuning of agents.
First time an AI model managed to write systeminstructions that I did not need to review a million times. First time a Model chose the right AI settings for tasks without me needing to intervene all the time. I created sophisticated subagents and skills so claude would do an alright job at it, but GLM, even without my harness, built things better than claude did with it.

For instance: classic rag will not do a great job for my type of use cases. I built on existing graph rag architectures added custom features and lots of complicated domain specific logics to get my AI systems to have better results than any of the competition.
For half a year I have now been working on a completely new retrival logic, trained custom embedding models for it, basically created a new type of rag, can't tell too much about it yet.
I let GLM work on it in a new branch, it gave insane suggestions we had not considered yet, also no other model did, idk how... and it implemented them in such a simple but yet brilliant way, upped our own benchmarks by 13% which is insane.

So all I can say: give GLM a try, you won't fucking regret it. Especially if you put it in the cursor harness, worked better than claude code.

[Felix Thöni @f_thoeniX]

@kimmonismus wondering if you might have any intel on https://t.co/B4op8iVJdT
I've heard from some sources that claim that they are giving correct numbers, but nothing I trust.

[Felix Thöni @f_thoeniX]

There is something fascinating about this number:
- the Chinese are mostly Chinese nationals / diaspora, that studies in the US (60% and EU 40%)
- the US are 75% who studied in europe 15% in India, 10% in China
- the EU are about 70% US (but other europeans) rest split between china and india
---
Overall most people big in ML studied in Europe.
Yet Europs best AI lab is Mistal who managed to bring out their new model mistral 3.5 medium in april 128B - that is as good as Gemma 4 with only 30B released in the same week... this is just so damn sad. they're litterally 1 year behind China in development. and roughly 1 1/2 behind the US...

[Felix Thöni @f_thoeniX]

@kimmonismus just finished continued pretraining today of GLM 5.1 on German law. Happy to have been betting on the right horse.
btw outperforms what Fable did by 40% on our internal evals. and Fable already outperformed the now 3rd place by more than tripple.

[Felix Thöni @f_thoeniX]

The sad thing is: It was and most likely will only be a dream.

I am certain, behind the scenes China is trying to get Anthropic come over to HongKong or something.

Meanwhile we here in Europe are awaiting doom due to our politicians.

Regulations everywhere and the worst thing: loopholes everywhere too, but you can only afford to know about them or use them when you have enough money.
Looking at what my tax advisor is doing - this should not be a thing - and it can only be done once the ammount of money is worth it, but then, holy shit.

The kind of politicians Europe needs are people who come from business and do not want the fucking job - those are the ones who know the system and could fix it. The ones that want the jobs, are usually clueless morons that are being tricked by the people who benefit the most from the system.

Problem is: can't force them and should not be able to. I would not wanna enter politics for a billion. That money is easier made in other ways nowadays.

[Felix Thöni @f_thoeniX]

I am on a serious Fable withdrawl. I really hope that they can figure this out.
I had a dream today that the EU-politicians suddenly stopped being dumbass idiots that fuck our future and offered Anthropic taxfree status for 10 years and lots of other benefits to come over to Europe...
That's what should be done now. Companies do not care about the countries they are located in. The US is no longer an EU Ally - we need to fight for US companies missaligned with the US government to change their seat of operations to Europe and deregulate.

MAAAN... I cannot think straight... I need Fable... My productivity dropped 80% compared to last week...

[Felix Thöni @f_thoeniX]

@OfficialLoganK In order to live in the present I need a new god model please. I am having Fable-withdrawl syndrome. No current coding model is satisfying to use any more. Please Logan, I believe in you, help me, gemini-3.5-pro?🥲

[Felix Thöni @f_thoeniX]

@LexnLin if anyone reaches AGI it will be Google. They're going in all directions, I am most impressed by their diffusion Gemma model. Insane speed and quite good still. - If they could scale this it would beat fable and all other models in coding in a heartbeat.

[Felix Thöni @f_thoeniX]

@ChrissGPT Claude 3 and Gemini 1.3 😂🤣 dayumn that's some sick ass new models

[Felix Thöni @f_thoeniX]

Some Info about European frontier models.
I am not allowed to disclose a lot, but there is a very massive german enterprise that is highly interested in starting a proper EU lab that will be able to create frontier models to compete with OpenAI and Anthropic. (Not the teny tiny Mistral models).
We have all the knowledge for it in Europe - the German open source AI scene has people that could very well work at anthropic.
A lot of lobbying is being done for this at the moment, to change regulations in our favour. I also know that the money is there to "pull a grok" and just suddenly be there among the top models - even release the top model at some point.
I have attempted in the past to get some money for a project at that scope going, but did not have the contacts necessary at the time. Now I do and I have been contacted regarding the project. I wish I could say more, I hope the things being discussed in the background will lead somewhere.
I for my part am confident that we will see a European Model be among the top 10 models in the world in 2027.
We have the brain power to create something crazy (more than the US), we have backing by massive companies, all we need is the political backing. This if anything helped our cause to get a huge EU-lab going.

[Felix Thöni @f_thoeniX]

@theo I love how 95% of commenters did not read your claude MD and now believe Fable answers like that 🥲

[Felix Thöni @f_thoeniX]

@OfficialLoganK @gatien7 Gemini 3.5 pro soon pls! ❤️

[Felix Thöni @f_thoeniX]

@mcnultydigital @Pontifex Funny question, but probably legitimate in these times. FYI I have a real human wife and 4 kids

[Felix Thöni @f_thoeniX]

@kimmonismus may I ask:
Are you using Codex as a general agent or for programming? I'm curious as I am still mostly using Cursor with Composer 2.5 for things that gotta run fast, and Claude Code CLI for programming harder stuff, and now running hundrets of millions of tokens per day with ultracode in claude code.

[Felix Thöni @f_thoeniX]

@kimmonismus Also I expext to pay 9.000€ to anthropic in token costs this month, for my coding tasks alone - and I guess in total this month will be the first my apps will cross 100.000€ in API costs. Insane thinking about that, seems surreal

[Felix Thöni @f_thoeniX]

@kimmonismus @kimmonismus you have to try out the new dynamic workflows in claude code. That is the real power of Opus 4.8. It's out for like 4 hours now and I have already burned 300 mil tokens and implemented 3 new features that would have taken me 4-5 days before. this thing is insane.