Shalva Dvaladze

Technology & Business Leader | 21+ Years in IT | Driving Innovation with Al | Building Complex Intelligent Systems | MBA, University of Sheffield

[Shalva Dvaladze @sdvaladze]

@JensenHuang Don't give up! You've never given up in 30+ years - keep going! Nvidia and the entire AI community still need your energy, courage, and support! 💚🖤

[Shalva Dvaladze @sdvaladze]

GTC Berlin 🇩🇪 | Two days of conversations, connections, and confirmation. 🚀
Day 2 was packed with meetings, valuable discussions, and opportunities to exchange knowledge with people shaping the future of technology.
For us at NIMBUS21, GTC was more than an opportunity to explore new developments. It was a valuable confirmation that the technologies we build and the methods we implement are at the forefront of an industry that continues to evolve at incredible speed.
We exchanged insights, strengthened existing relationships, established promising new connections, and explored perspectives that will help us continue moving forward.
One of the most valuable aspects of events like GTC is the opportunity to compare ideas, challenge assumptions, and see how your work aligns with the direction the industry is heading.
We leave GTC Berlin with new connections, valuable insights, and even greater confidence in the direction we're taking at NIMBUS21.
A genuinely fruitful experience — and plenty to build on from here.
Thank you to everyone we met and connected with throughout the event. Looking forward to continuing the conversations.
#GTCBerlin #NIMBUS21 #AI #Innovation #Technology

[Shalva Dvaladze @sdvaladze]

Day 1 at GTC Berlin 🇩🇪 | The energy around AI is incredible. 🚀

Day 1 started inside Tempodrom, where, thanks to our partners, we had the opportunity to experience the event from a great spot in the room.

Beyond that, I explored and participated in different parts of the event, gaining new perspectives on the rapidly evolving AI ecosystem.

What stood out to me wasn’t just the scale of GTC Berlin, but the energy, ideas, and conversations bringing people together around the future of AI.

Sometimes, the most valuable part of an event isn’t just what happens on stage. It’s being present, listening, exchanging ideas, and connecting with people who are building what’s next.

Day 1: Done. Day 2: Let’s go. 🔥

Looking forward to more conversations, new perspectives, and insights worth sharing.

If you’re at GTC Berlin, what has been your biggest takeaway so far?

#GTCBerlin #AI #ArtificialIntelligence #Innovation #NIMBUS21 @nvidia @NVIDIAGTC @JensenHuang

[Shalva Dvaladze @sdvaladze]

GTC BERLIN - dm if you are there too lets connect .

[NVIDIA AI @NVIDIAAI]

Berlin is an epic location for @NVIDIAGTC!

Follow along for announcements from Jensen's keynote tomorrow.

[Shalva Dvaladze @sdvaladze]

@benedictk__ Nvidia ?

[Shalva Dvaladze @sdvaladze]

@trq212 talking about slowing down . opus 5.5 is very slow now is that intentional ?

[Shalva Dvaladze @sdvaladze]

GTC Berlin is on my calendar this October! I'll be exploring the latest innovations in AI, accelerated computing, and robotics, and GTC is where a lot of what's next is shown for the first time. If you're going to be there, let's connect in person. Drop a comment or send me a message.

https://t.co/YUeJJvdqpE

[Shalva Dvaladze @sdvaladze]

First impressions with Muse- Interesting !
muse-wonder-1.0 · xhigh - 👏
@Muse

[Shalva Dvaladze @sdvaladze]

hey @Muse,
I'm getting error when i want to subscribe.
Is there any way to subscribe ?

[Shalva Dvaladze @sdvaladze]

@finkd @AIatMeta @finkd - great job , lets see how it performs in real life. Will get back with results.
BTW : ad video is just amazing ! Love it .

[Shalva Dvaladze @sdvaladze]

@giadotai @thecsguy That is where your 4K should go -> “use lots of tokens”😅

[Shalva Dvaladze @sdvaladze]

I don’t think statelessness means the full tool catalog has to be sent on every request. That feels more like a harness design limitation. (Most harnesses solved that long ago)

A harness could expose only a few stable meta-tools: search_tools, mount_tool, unmount_tool.

The actual MCP registry lives outside the model context. The LLM sees a compact, session-scoped capability index, searches it when needed, then the harness resolves the MCP server and exposes only the relevant tools.

So instead of:
10,000 tools → prompt

you get:
intent → search → resolve MCP → mount relevant tools → use → unload

And if changing the tool schema hurts prompt caching, keep the model-facing interface stable and route execution through a universal tool proxy.

I think this is primarily an orchestration/harness problem, not a protocol problem.

[Shalva Dvaladze @sdvaladze]

Maybe we don’t need another agent protocol.

We need self-mounting MCP.

Agent needs a capability → discovers MCP → checks trust/auth → mounts → uses → unloads.

MCP is already the common interface. The innovation may be orchestration, not another protocol.

What am I missing?

[Shalva Dvaladze @sdvaladze]

@SebastianRoehl Just to compare each of our team members have minimum 60+ machines in cloud developers 100+. Minimum spec for machine is 8 CPU cores 16 GB RAM .

[Shalva Dvaladze @sdvaladze]

When you have this brilliant idea to invest 1 week to build an agent which reads emails and buys you coffee - please be informed that, others also have same ideas, same coding agents and are already doing exactly same maybe a little more different - brand colors .
Solve the problem which others can’t. Be creative you can do better then that !

[Shalva Dvaladze @sdvaladze]

@GoogleDeepMind Ja , better to compare with @cohere or @VoyageAI

[Shalva Dvaladze @sdvaladze]

The biggest bottleneck to AGI isn’t the model.

It’s memory.

We evaluated 10 of the leading AI memory systems.

Different architectures. Different marketing. But fundamentally, most are still playing on the same field.

No real breakthrough yet.

At NIMBUS21, we’ve been researching this problem from a different angle - and the results are getting interesting.

Too early to share the details.

But one thing is becoming clear:

The real AI race may not be won by whoever builds the biggest model.

It may be won by whoever solves memory.

Keep an eye on NIMBUS21 AI. 👀

[Shalva Dvaladze @sdvaladze]

@hwchase17 We dont need agents which sleep at night 24/7 work that is true value ))