Georg Unterholzner

Chief of Yelling at Agents @Dynatrace. Data scientist turned software engineer.

[Georg Unterholzner @georg_dev]

rate my setup :)

[dax @thdxr]

things would be really crazy if deepseek was an american company

[Georg Unterholzner @georg_dev]

Random guy in the tram next to me just had a phone call asking someone how he can buy Anthropic stock.

We're approaching peak bubble.

[Georg Unterholzner @georg_dev]

@adiccit You switch to deepseek

[Georg Unterholzner @georg_dev]

Sorry to break it to you but #̢̛ͥ̀̽̽ͨ̌̓͆̇̀̽ͦͨͫͯ̏̎̇̈̚͜҉̕͢͏͡҉̸̶̛͢͡҉̵̨̭̳̘̹̫͔͖̜̠͖̫̲̖̱̤̻̖͉͢͞͡ͅͅ͏̸̷̢͑̊̀̐ͣͭͭ̉͂ͨ̽͒̊̎͛ͪ̊̎ͩͧ̀͝͠͠͏̸̵̵̷̵

[Georg Unterholzner @georg_dev]

oh god please let the bubble pop before those idiots start banning all open weights models

[Zack Korman @ZackKorman]

The order banning DeepSeek in the US government is absurd. Why ban US-hosted DeepSeek? "The models themselves may contain malware".

The proof? The DeepSeek iOS app has vulnerabilities.

[Georg Unterholzner @georg_dev]

DeepSeek V4.1 Flash 1017 is a really nice model. Having lots of fun with it so far.

[Georg Unterholzner @georg_dev]

@ozandagdeviren Haha of course there's a product for that

[Georg Unterholzner @georg_dev]

Shower thought: if you hire for an engineering position, why not give the interviewee a small task to solve w/ AI and ask them to not only hand in the solution, but also their full Claude conversation history?

Think about how much you could learn from their transcript: ...

[Georg Unterholzner @georg_dev]

@HeyGarrison "I'm not gonna read that all. Just do what you think is right"

[Georg Unterholzner @georg_dev]

- thinking patterns
- problem understanding
- AI proficiency
- engineering knowledge
- delivery speed
- ...

There's so much information in the transcripts. You would even learn how to improve your interview assignments.

Why not?

[Georg Unterholzner @georg_dev]

@thorstenball Depends on the complexity of the system. For >95% of projects, that's probably true.

However, there are systems in which your agent yoloing stuff onto main will lead to serious consequences later down the line and you better do your due diligence before merging anything.

[Georg Unterholzner @georg_dev]

@simonbs I think humanlayer supports on prem in their Enterprise plan. Haven't tried it out myself, but looks promising imo.

[Georg Unterholzner @georg_dev]

@scott_stouffer That's correct, Claude.

[Georg Unterholzner @georg_dev]

I really despise how anglocentric Claude is. All the time: you send Opus 5.5 to gather some multilingual data. It *filters for English* and reads only that.

Bro, don't you speak 100+ languages or what? Aren't there internal evaluations for multilingual retrieval at Anthropic?

[Georg Unterholzner @georg_dev]

@wertform_mcm Kannst easy 10 Jahre rauszögern. Bring ihn einfach auf die Idee, es vorher mit SAP zu verknüpfen.

[Georg Unterholzner @georg_dev]

I can't imagine they're not planning to open this at least to selected partners at some point (think: uber, airbnb, etc.).

It will be small self contained web apps in iframes. That part is easy.

The challenging part is conceptual: how much should the outer app (= chatgpt) be aware of what happens inside the iframe?

Feels kinda frustrating to me, if my AI assistant shows me a widget, but then has no clue what happens inside it. Moreover, I doubt that non technical users would intuitively understand that there is a conceptual border between chat and widget. On the other hand, if you relay information from the widget back to the LLM, how do you cache that?

[Rabi Shanker Guha @rabi_guha]

ChatGPT launched their version of Generative UI yesterday. We took it apart

tldr; the agent never write raw code. It emits DIL (a new language), the server compiles and streams it, the client renders it using a component set so that it feels native https://t.co/EzfM43fgkf

[Georg Unterholzner @georg_dev]

@KolumbusKc @mattpocockuk at least you have funny stories to tell

[Georg Unterholzner @georg_dev]

@diegohaz You can configure the subagents cache TTL in the settings or via env var: https://t.co/0bBQrij7rq

Gotta do this now as 5 min is way too low.

[Klaas @forgebitz]

honestly, most code before ai was also slop

[Georg Unterholzner @georg_dev]

@rahulvrane @GeoffreyHuntley Asking for my wife who's a pharmaceutical chemist and gets flagged constantly for literally trying to develop a drug against a disease w/o a cure: can you work properly with grok, or is it also refusing to work but just less often?