Nenad Mancevic

Principal Engineer @HBOMax. Building petabyte-scale multimodal AI. Video intelligence, semantic search, model orchestration, agent-callable infra. Views my own.

[Nenad Mancevic @manca1]

This sounds like a great opportunity! Cloudflare is one of those companies you can’t not love. I’ve been a happy customer for over a decade… the amount of innovation the team is capable to ship and keep on improving the core products is second to none! Kudos to @eastdakota for making what Claudflare is today!

[Nenad Mancevic @manca1]

@jxnlco Why do you have “End conversation” button? Use it when the robot gets out of control, huh? :)

[Nenad Mancevic @manca1]

@dhh Yet they keep releasing new macbooks? Not sure what the play was with MacBook Neo, but my 7yo wants it, despite having an iPad. Maybe Apple will eventually move to iOS all across devices… the new Neo already uses the A18 chip.

[Nenad Mancevic @manca1]

@sundeep Exactly! Grind your way to obscurity and never touch grass.

[Nenad Mancevic @manca1]

@samlambert Simple is not easy. Always reminds me of that great Rich Hickey’s talk.

[Nenad Mancevic @manca1]

@ivanburazin I agree with the idea, but it is much easier said than done. Any post-training or RL technique, requires not just deep expertise in the model architecture, but also very high quality hand curated data. You can’t just slap QLoRA or GRPO and hope for the best.

[Nenad Mancevic @manca1]

@GergelyOrosz With all the new shiny tools available to us, it’s veryyyy tempting to go out and try to make it big by starting your own thing. Asking claude code or codex to build you the next pick-your-favorite-saas clone is not exactly “pushing the boundaries of AI”, though.

[Nenad Mancevic @manca1]

@kmad This is great. Would be good to add the total token spend on these tasks so it can be compared to traditional CV models both on perf and cost.

[Nenad Mancevic @manca1]

@_joe_harris_ But, it’s hard :) Seriously, though, I agree that we’ve just scratched surface on what’s possible in the robotics field with the new AI capabilities including the SOTA physical world models and VLMs. Super exciting field!

[Nenad Mancevic @manca1]

@brendanrc2 always be deploying!

[Nenad Mancevic @manca1]

@HeapSmasher Yeah. I feel like people are narrowly focused on solving just the agent isolation, often forgetting the environment within which the agent actually runs.

[Nenad Mancevic @manca1]

There’s been a lot of discussion about the proper agent sandboxing and how to avoid agents “escaping” the sandbox and hacking other sites, or worse share your private keys and secrets outside.

My take is that the container is not the sandbox. The environment is. A thousand agents with perfect per-agent isolation still share the package registry, DNS, and the secrets vault, and that shared layer is where escapes actually happen.

Instead, focus on building a secure cloud environment (a separate Google project or AWS account dedicated to agents) with strong network isolation (secure, firewalled VPC perimeter), a private package registry and a single egress gateway which only allows access to outside inference endpoints.

Once you solve the isolation at the environment level you can safely run agents on any compute inside it: VMs, containers, serverless, etc.

[Nenad Mancevic @manca1]

@skalskip92 Do you plan to self-host it once they drop the weights or use it via an API? I guess hosting a 2.4T parameters model won’t be an easy task, though.

[Nenad Mancevic @manca1]

@sundarpichai @JeffDean @Sanjay_Ghemawat @Google Wow. The end of an era. Amazing run, Jeff! Good luck with @DiscoLoopAI.

[Nenad Mancevic @manca1]

@0xblacklight I am also trying to understand what is so hard about building a sandbox environment in which you can run compute, have strong network isolation, constrained set of permissions, etc.

https://t.co/u5MUeAN070

[Nenad Mancevic @manca1]

@rakyll Build a cloud sandbox environment with strong network isolation (secure VPC perimeter), data control, logs/audit, etc. and then run the agents in whatever compute you’d like (VMs, containers, serverless, etc).

[Nenad Mancevic @manca1]

@0xblacklight For sure. It’s just much easier for regular folks to fire off a prompt or two to chatgpt or google AI, get the response and move on. I feel we are still far from significant agent adoption in regular population… maybe the humanoid robots will help.

[Nenad Mancevic @manca1]

@fayazara @eastdakota Congrats on the launch! Curious how do you handle agents identity and cross-system auth?

[Nenad Mancevic @manca1]

I'll just leave it as is. What a year...

[Andrew Curran @AndrewCurran_]

At first I thought Bloomberg forgot to add DeepSeek's pricing to the chart.

[Nenad Mancevic @manca1]

@realmadhuguru > Learn what users actually want and iterate.

This is the hard problem that few get right. That's why most are stuck there...