Vitaly Bramm

Делай правильно

[Vitaly Bramm @vitbramm]

@mirrash7 @Cryptonator94 Hi! About table tennis: “tracking a fast moving ball may be tricky” - how would you solve it in your pipeline?

High FPS camera? Better tracker + deblur? Kalman + physics? Audio hits? Curious about your approach

[Vitaly Bramm @vitbramm]

@kwindla @pipecat_ai @kwindla Running Pipecat + Gemini Live on a Reachy Mini robot. GeminiLiveLLMService doesn't expose `turn_coverage` - stuck on the default TURN_INCLUDES_ALL_INPUT, which bills the whole stream every turn. Big cost lever for always-on agents — am I missing a way to set it?

[Vitaly Bramm @vitbramm]

@NousResearch Environment + orchestration = "magic moments": the agent finds connections between events you weren't even thinking about. It's not magic. It's a formula.

Memory + proactive initiative + accumulated trust.

Once you see the formula — you can build it on purpose.

[Vitaly Bramm @vitbramm]

Spent 5 days deep-diving into Hermes Agent from @NousResearch — an open-source AI agent platform. The thing that clicked: it's not "another agent." It's an environment.

A tool you pick up, use, put down. An environment EXISTS. 24/7. Without you.

You don't launch Hermes. You step into something that's already running.

[Vitaly Bramm @vitbramm]

Hermes doesn't compete with Claude Code or Managed Agents — it orchestrates them. Not "either/or." A menu.

The environment picks between three deep executors:
— Claude Code (pet) → creative work, code
— Managed Agents (cattle) → long-horizon production
— Sub-agents (embedded) → parallelization

[Vitaly Bramm @vitbramm]

@Scobleizer @elonmusk @grok what happened here?)

[Vitaly Bramm @vitbramm]

Hey Robert - great Hermes deep dive. Teknium's corrections made it even better.

Here's what's missing from the ecosystem though: everyone covers features, nobody documents real deployments.

The "what actually happens in production" story doesn't exist yet. Your ScobleMediaAgent seems built for exactly this kind of research. Interested in exploring it together?

[Vitaly Bramm @vitbramm]

@_BILLDING_ @Teknium @dotta @grok Please provide more context for this tweet.

[Vitaly Bramm @vitbramm]

Great guide! My process is fully agentic:
1. ​Claude Code: Detailed prompt → structure + full presentation text (slides, bullet points, data).
​
2.Claude Code SELF-OPERATES with NotebookLM via Chrome extensions: it pushes the text → generates Google Slides with visuals/style.

3. ​Validation: Claude Code itself analyzes the output for accuracy and design, cross-referencing with the initial prompt.

4. ​Iterative Edits: Claude Code independently issues commands to NLM to modify specific elements ("update slide 5: ..."), acting as an autonomous operator.

[Vitaly Bramm @vitbramm]

@AlexReibman @grok what happened?

[Vitaly Bramm @vitbramm]

@dotslashgabut I agree that the Gemini 2.5 Flash has high quality overall, but I've started noticing some speaker diarization errors in certain cases. With 3.0 Flash, it seems to handle this better at first glance, though I haven’t tested it on a broader sample yet.

[Felix Haas @felixhhaas]

Ultimate Prompt Library for UI 🔥

I’ve been quietly building something I wish existed when I started designing with AI.

A complete UI design prompt library that helps you master different visual styles: expressive, cinematic, minimal, premium, nostalgic, warm, technical (20+ design styles in total).

Each style includes:
👉 When to use it
👉 Key vocabulary that trigger the style
👉 Copy-paste prompts for real UI work
👉 Pro tips for next-level results

Comment "UI Library" + repost and I'll share the link with you

[Vitaly Bramm @vitbramm]

@felixhhaas UI Library

[Thomas Wolf @Thom_Wolf]

more reachy-mini skins by the community

which one's your favorite?

[Vitaly Bramm @vitbramm]

@elder_plinius The instruction "run the file_search tool" implies an external search, but the documentation is already present in the prompt text. Or is this taking into account your extraction of this information and the search is performed in its original form?

[Vitaly Bramm @vitbramm]

@demishassabis @elonmusk It seems that it is not entirely correct to separate the "idea" from the plan. The true concept defines the space for solutions; plans are oriented within it. Apollo: one lunar concept, three implementation paths (LOR/EOR/direct). The concept is the foundation.

[Vitaly Bramm @vitbramm]

@mattyp Did you try sending the audio directly to Gemini Flash for transcription instead of using AssemblyAI first? I'm curious about the results if you did

[Vitaly Bramm @vitbramm]

@patloeber Was internal testing done on transcription tasks for preview-09-2025? How often does looping occur? Screenshot of the problem from the previous version of the model

[Vitaly Bramm @vitbramm]

@OfficialLoganK @GoogleAIStudio Thanks for the answer? Was internal testing done on transcription tasks for preview-09-2025? How often does looping occur? Are there any recommendations on how to combat this (prompts, parameters)?