Dennis Pettas 🦊

Building things that think | Surfing on/off-chain
Before data @binance, @glassnode, @arkham, @optimism
All neural outputs are my own or my dog's

[Dennis Pettas 🦊 @dioptx]

What’s the equivalent of this in real life?

[Dennis Pettas 🦊 @dioptx]

@cobaltdigital33 Literally closing the laptop and touching grass equals less net work when this happens.

[Dennis Pettas 🦊 @dioptx]

@nithin_k_anil @DrAIExpert the tell within the tell is which surface they shipped first. analytics, not ads, not search console. read-heavy query workloads. that tells you exactly where MCP has proven out and where the gaps still are.

[Dennis Pettas 🦊 @dioptx]

Just published a deep dive on why AI reasoning has no addresses.
Every other system you debug has coordinates. Stack traces, span IDs, flame graphs. AI reasoning? You re-read an essay and hope the bug jumps out.

I built the fix. Here's what it looks like in practice.

🔗 https://t.co/GOLmbqRQ8E

[Dennis Pettas 🦊 @dioptx]

https://t.co/oKWFTTeLtL

[Dennis Pettas 🦊 @dioptx]

https://t.co/oKWFTTeLtL

[Dennis Pettas 🦊 @dioptx]

@freeCodeCamp Shipped a couple of MCP servers to npm & can confirm the gap is real.

No tutorial covers what happens when 3 agents hit the same state file or how schema validation fails silently. We need more of this.

[Dennis Pettas 🦊 @dioptx]

@CryptoCred Agentic trading will become the norm to some degree. Agents that predict their failures will be the real winners, not humans.

[Dennis Pettas 🦊 @dioptx]

@enesakar Running the MCP version daily in Claude Code and it's a real token win. CLI closes the gap for non-MCP setups.

[Dennis Pettas 🦊 @dioptx]

@stacy_muur Yield and rug risk are the same variable. You asked for
the highest and lowest value of the same number.

[Dennis Pettas 🦊 @dioptx]

@levelsio Half of the people saying MCP is dead are now putting tool lists in their Claude.md. That is a discovery manifest. It is less standardized MCP.

[Dennis Pettas 🦊 @dioptx]

@rafalwilinski Correct for remote MCPs. Most deployments today are local stdio with zero auth in play. The OAuth case is real but not the majority use case yet. Yes, new paradigm be "shakey", but that doesn't mean we won't converge to something useful.

[Dennis Pettas 🦊 @dioptx]

@TheGeorgePu Not just a wrapper. The distribution moat is real short-term. The question is whether Anthropic and OpenAI keep tolerating that margin compression.

[Dennis Pettas 🦊 @dioptx]

Turns out building things is inversely correlated with posting about building things.

[Dennis Pettas 🦊 @dioptx]

@Defi_Warhol @DefiLlama @artemis @nansen_ai The gap is exchange-level intelligence: @cryptoquant_com for reserve flows + miner capitulation, @glassnode for long/short-term holder behavior. Macro timing signals TVL dashboards don't cover.

[Dennis Pettas 🦊 @dioptx]

@nicbstme @firecrawl @p0 CLI vs MCP conflates two problems: discovery (what tools exist?) vs invocation efficiency. CLI wins on efficiency. MCP wins on discovery. The bottleneck depends on your use case.

[Dennis Pettas 🦊 @dioptx]

@harshbhatt7585 @karpathy Reward specification is the bottleneck here, not the RL algo. Getting incremental returns in 24 experiments means the env modeling is actually decent and that's harder to build than the agent itself.

[Dennis Pettas 🦊 @dioptx]

@Chris_Worsey @karpathy Killing the worst-Sharpe agent without regime weighting is the wrong selection pressure. Under drawdown, high-Sharpe agents look terrible. You're pruning the ones right about regime shifts.

[Dennis Pettas 🦊 @dioptx]

@const_reborn Your constraint isn't compute cost, it's signal quality. Hyperliquid's order book is unusually clean. That edge disappears if the data layer degrades.

[Dennis Pettas 🦊 @dioptx]

@aakashgupta Perplexity hosts an MCP server so Claude Desktop and Cursor can discover them. Internally, they run tight APIs which is rational for a production stack at $656M ARR. Two separate problems: developer discovery vs. internal execution. The CTO spoke to one of them.

[Dennis Pettas 🦊 @dioptx]

@jinglejamOP 🔴