Bernd Holbein

Accelerating Digital, Data & AI Strategies for Executives | Former Google Leader, CTO and Board Member | Entrepreneur

[Bernd Holbein @berndholbein]

Joined a start-up as a CTO/CTO after 8 years at Google to enable people live a more balanced life and break with some outdated e-commerce conventions #umcomfortablyexcited #startup #healthcare #WishMeLuck

[Bernd Holbein @berndholbein]

Who would have thought that AI generates new jobs this fast.

#ai #aiexpert #cheetah

[Bernd Holbein @berndholbein]

5️⃣ Mixtral 8x22B - Stands out for its cost-performance ratio. Offers function calling and constrained output, making it a versatile and budget-friendly option.

Which of these LLMs have you tried, and what is your experience ?

#AI #LLM #MachineLearning #FoundationModels

[Bernd Holbein @berndholbein]

4️⃣ Gemini 1.5 Pro by Google - Breaks records with an incredible 1M token context window and impressive MMLU performance. Speedy and more affordable than GPT-4 and Claude 3 at $21 per 1M tokens. The catch? Not available worldwide yet (e.g. Germany still waiting).

[Bernd Holbein @berndholbein]

🥉 LLama 3 (70B) - The new open-source powerhouse. Lightning-fast throughput and excellent text understanding at a fraction of the cost compared to the top 2. Available through Groq or self-hosting, with fine-tuned versions capable of function calling.

[Bernd Holbein @berndholbein]

🥈 GPT-4 Turbo - A close contender with a performance just shy of Claude 3 Opus, especially with complex prompts. Still on the pricier side at $30 per 1M tokens, but a solid choice for quality results.

[Bernd Holbein @berndholbein]

🥇 Claude 3 Opus - Takes the crown with the highest MMLU score, boasting a massive 200k context window for superior text comprehension and generation. However, it's a premium choice at $75 per 1M tokens. Good news: API access is now open

[Bernd Holbein @berndholbein]

I tested several LLMs, so you don't have to.

Buckle up:

Here is my personal current Top 5: ⬇️

[Bernd Holbein @berndholbein]

Artificial Intelligence is like teenage s...
everyone talks about it,
nobody really knows how to do it,
everyone thinks everyone else is doing it,
so everyone claims they are doing it.

[Bernd Holbein @berndholbein]

Wie sind deine Erfahrungen mit Remote Meetings? Und welche Tipps hast du für bessere Remote Meetings?

[Bernd Holbein @berndholbein]

Remote Meetings sind Mist!
Persönliche Meetings sind oft schon Zeitverschwendung, aber Remote...

Das Problem:
- Ständige Ablenkung durch Mails, Slack, Tabs
- Kaum visuelles Feedback oder aktives Zuhören
- Viel zu schnell und oft unüberlegt eingestellt

[Bernd Holbein @berndholbein]

Ansonsten gelten die Regeln für ordentliche Meetings.
- Püntklicher Start und Ende
- Klare Agenda und Follow-ups
- Nur relevante Personen im Call
- möglichst kurz

Ich arbeite seit 4 Jahren komplett Remote und liebe es. Gute Remote Meetings erfordern aber Disziplin.

[Bernd Holbein @berndholbein]

@gauntface 40 years, left Google after 8 years. Joined a startup as a CTO and can totally relate. I know nothing...

[James Singleton @ThisIsSingleton]

"In my experience, it can always be done — even if it can’t, it can." - @ScriptedAlchemy

I love it!

[Bernd Holbein @berndholbein]

If you’re using sGTM and get 400 http errors serving ec.js from your server domain, this is for you. @SimoAhava FYI

[Bernd Holbein @berndholbein]

@RickDronkers Thanks Rick. This drove me nuts. Thanks for the straight forward solution. Btw. As far as I can see this issue does not only impact ua related data. For me every subsequent hit after a ua event didn’t went through.

[Jared Spool @jmspool]

Great designers don’t fall in love with their solution.

Great designers fall in love with the problem.

[Bernd Holbein @berndholbein]

@ScriptedAlchemy @NetflixEng Awesome

[Tim Kietzmann @TimKietzmann]

When you sell AI but use regression in production.

[Zack Jackson @ScriptedAlchemy]

Kind of got next.js servers to hot reload federated modules that are streamed. If we can get it working, next apps could be “always on”, no need to restart or redeploy to get new code.