Gemini 3 is
- slow
- severely limited even in the paid plans
- feature limited
- outdated
Goog has a full vertical for AI, own hardware, cloud, development teams. They could outperform Anthropic within months.
It's just not profitable to do so now.
Personal account.
A helpful person in Information Security & Digital Forensics. A trader who minds market fundamentals. AI / ML dev. SDR nerd. Humor.
2 weeks retention.
| Develops | Python, Kotlin, Rust, JS |
| InfoSec's | Outside in / Inside out |
| Comliance's | Privacy, Environment |
| Hacks | SDR, ML, Ciphers, Forensics |
| Chats | https://wishi.keybase.pub/foobar.html |
Gemini 3 is
- slow
- severely limited even in the paid plans
- feature limited
- outdated
Goog has a full vertical for AI, own hardware, cloud, development teams. They could outperform Anthropic within months.
It's just not profitable to do so now.
Ich halte es für ausgeschlossen, dass der Grund für den Bahnausfall das GRM-R System ist.
GSM-R ist weder alt, noch ungeeignet. Das ist ein GSM in einem anderen Frequenzband. 5G macht auch nichts anderes, vom Prinzip her.
Software-Updates klingt jedoch so, als ob da jemand mit Claude Code Firmware gepusht hat. Das ist wahrscheinlicher.
It's astonishing that it's easier to get code from Opus or Sonnet than getting documentation for code.
It seems getting correct, verified and non-hallucinated documentation is an unsolved problem.
- invented functions
- mis-represented use cases
- useless information
With code you can reinforce train the model. With documentation you cannot easily do this. Text compression doesn't solve the knowledge compression issue, and compressed text isn't necessarily correct.
OpenUltraCode in action with opencode:
https://github.com/norandom/OpenUltraCode
1. structured assessment
2. step by step
3. multi agentic reasoning
Free. Any model. Any time.
I also put the theme somewhere, but I believe the Borland fanbase is rather small ;)
I am creating a Harnessx based security scanner for code.
https://github.com/norandom/OpenUltraSAST
https://arxiv.org/abs/2606.14249
1. scan Python, JavaScript, Java, C / C++ vibecode
2. let it find things
3. let it tune itself to eliminate false positives (and false negatives if possible)
4. run it with different models
So far it looks interesting enough. With LLMs you can extend the concept towards the security design. But I'd keep that separate.
I think Kubero can be useful to host vibecode projects at scale in a business.
https://github.com/kubero-dev/kubero
1. create a Skill how to deploy
2. put the rules in the Skill
3. deploy Kubero in a separate Kube cluster (NixOS k3s is a 3-liner to setup a Kube node / master)
4. put that on a Proxmox or similar (cheap) infrastructure
5. push some template repos
Companies are drowning in people with AI skills. What's needed are decision makers, who define how to scale this.
I just ran Claude code /goal on my personal factor invest (Python) framework over night.
It didn't get so far. I think developing agentic friendly frameworks, where they can run data-simulations over nights, is key.

I don't recommend any private investor (retail, non-professional) to buy any IPO / risky derivative.
If you are a professional (Quant, Portfolio Manager), you don't give recommendation anyways, and you don't hang out in social media chats to give people finance advise.

https://www.youtube.com/watch?v=aUvRZcPchNM
Or in Quant: do a reverse DCF and check the rDCF vs. last day's close.
Then do a 5y / 10y DCF model to get TV.
Then decide if he is correct.

https://www.youtube.com/watch?v=3TWODrn4aAE
Interesting perspective
