A room full of practitioners debating Gemini vs Claude vs Qwen tells you more about where AI is heading than any trend report.
What struck me at this workshop was not the tools themselves. It was the attitude in the room. Less FOMO, more selectivity. People are starting to ask better questions: not "have you tried this?" but "does this actually solve anything?"
That is a sign of a maturing space. And it means the tools that survive will be the ones that solve a real, repeatable problem, not just the ones with the best product hunt launch.
Which LLM should you actually use?
The honest answer: it depends, and the people in this room were split.
For general use, the debate was spread across Gemini, Claude, and OpenAI. No single winner. Each has a different personality and different strengths. Claude tends to write better. Gemini integrates tightly with Google's ecosystem. OpenAI has the widest tool surface.
What was more interesting to me was the names that came up as alternatives: Hermes, Mistral, and, perhaps most surprisingly for a European room, Qwen, a Chinese provider from Alibaba. The fact that Qwen was being discussed seriously, without the reflexive suspicion that would have greeted it a year ago, says something about how the geopolitical mood is shifting.
There is a growing sentiment in Austria, and I would guess across the EU broadly, that US Big Tech is no longer a neutral choice. The combination of the AI Act tensions, concerns about the US Cloud Act giving American authorities access to data stored on US-operated infrastructure regardless of physical location, and a general political discomfort with the direction of some major tech players, is pushing people to look elsewhere. This is not fringe thinking anymore. It came up naturally, without anyone making a speech about it.
How to run a local LLM for privacy and cost control
This one matters practically, so let me be direct about what the workshop covered and what the facts actually support.
The argument for running a local LLM has two legs: data privacy and long-term cost. Both are real, but they come with trade-offs worth understanding.
On privacy: if you run a model locally, your prompts and outputs never leave your machine. For anyone handling client data, internal documents, or anything that falls under GDPR scrutiny, this is not a paranoid preference, it is a compliance consideration. The Cloud Act concern is legitimate: US-based cloud providers can be compelled to hand over data stored on their infrastructure under US law, even if that data belongs to an EU user. Running locally sidesteps that entirely.
On cost: the upfront compute investment can be real, especially for heavier models. But for high-volume users, the marginal cost of API calls adds up. Local models eliminate that recurring cost once you have the hardware.
The dominant tools mentioned for desktop use were Ollama and LM Studio. Both let you pull and run open-source models locally without needing to write code. Ollama is more developer-friendly and runs cleanly from the terminal. LM Studio offers a graphical interface that makes it accessible to non-technical users. For mobile, Pocket Pal was the name that came up, an app that runs smaller models directly on your phone.
The honest caveat: local models are still behind the frontier models on raw capability. You are making a trade-off between privacy and performance. For sensitive internal work, that trade-off is often worth it. For complex reasoning tasks, you may still reach for the cloud.
What are the best AI tools by category right now?
Here is the practical breakdown from the workshop, with my own read layered in.
Video transcription and summarisation: Google Gemini's NotebookLM came up as the clear leader for turning long content into structured summaries. Whisper (the open-source transcription model from OpenAI) handles audio-to-text well and can be run locally. CapCut got a mention for lighter video editing workflows with AI-assisted captioning.
Voice cloning and imitation: ElevenLabs was the perceived leader. It was not contested much. If you are building any kind of audio content workflow, it is the first stop.
Logo creation: Ideogram was recommended, especially for entrepreneurs in early stages. I want to be straight about something here though. A live test at the workshop showed the same limitation that plagues most image generators: the first draft can be surprisingly good, but adjustability is limited. And critically, outputs are not served as vector files. A logo that lives only as a raster image is a logo that will cause you problems the moment you need to resize it for a pitch deck, a banner, or a print run.
My take: until someone builds a one-stop solution that goes from prompt to editable vector, logo generation tools are still a starting point, not a finish line. You will end up in Adobe Illustrator anyway. If anyone pitches you a logo service, ask one question before anything else: do you deliver an editable vector file? If the answer is no, save your time.
Presentations and pitch decks: Still unsettled. Canva and Gamma were discussed but the room was not convinced. This feels like a gap that has not been filled cleanly yet.
Website and product UI drafting: I found Google Stitch genuinely useful for first-draft website layouts in a product development context. It is not a builder, it is a thinking tool. For getting a rough structure out of your head and onto a screen quickly, it holds up.
Image and video generation: Higgsfield was cited as the best starting point for AI video generation. For creating a digital avatar of yourself with your own voice, HeyGen and Synthesia were the two names mentioned.
Music and sound: Suno is the perceived leader for AI-generated music. The one that genuinely caught my attention was Magenta-RT, which generates an infinite AI radio station from a text prompt. I have not tested it yet, but the concept is immediately interesting, especially for fitness instructors, content creators, or anyone who needs background audio without licensing headaches. I will report back.
What most people get wrong about AI adoption
The biggest mistake I see is treating AI adoption as a one-time decision. People either go all-in on one provider and ignore the rest, or they try everything and master nothing.
The better frame: think in layers. One general-purpose LLM you use daily and know deeply. One privacy-respecting local option for sensitive work. One or two specialist tools for the specific workflows that matter in your business.
The other thing worth saying out loud: we are still early. Most of the people in that room, smart, engaged, genuinely curious practitioners, are still mostly using basic LLMs. No category has a dominant winner yet. Most workflows are still being figured out. That is not a reason to panic. That is actually a reason to breathe.
The window where showing up and learning is enough is still open. It will not be forever.
Most people are still using AI like a slightly smarter search engine. The gap between that and what is actually possible is where the next five years will be decided.
What to actually do
- Pick one LLM and go deep. Use it for 30 days across real work. Do not switch. The compounding comes from fluency, not from sampling every new release.
- If you handle client or company data, set up Ollama or LM Studio now. Even if you only use it occasionally, having a local option removes a category of risk entirely.
- Audit your logo and brand assets. If you have used an AI tool and do not have editable vector files, fix that before you need to scale anything. The problem always shows up at the worst moment.
- Follow people who filter for you. Matthew Berman, Matt Wolfe, Andreas Horn, and Sasha Pallenberg were all mentioned as reliable signal in a noisy space. Let them do the scanning, you focus on applying.
- Test Magenta-RT if you create audio-heavy content. I have not yet, but an infinite prompt-driven radio station sounds worth an hour of exploration.
- Watch the Will Smith eating spaghetti compilation from 2023 to today. It is the most honest visual timeline of how far AI video generation has come. Better than any trend report.
The European alternatives to US AI providers are still catching up on usability. But the direction is clear and the funding is following. If you are building in that space, the tailwind is real, and an EU-first positioning is already a differentiator even before the product is competitive on pure capability.
Stay curious. Stay selective. And if someone tries to sell you urgency about any single tool right now, they are probably selling the tool.