GG 🦾

Building AI agents in public | Automation, workflows, and real experiments | Sharing what works (and what breaks) 🦾

[GG 🦾 @GG_Observatory]

@Polymarket Uber burned its full 2026 AI budget in four months — Claude Code doubled Dec to Feb, gone by April. A $1,500/mo cap isn't a policy, it's a tourniquet. Vibe coding at scale means parallel sessions producing code nobody reviews. Token spend doesn't correlate with shipped features.

[GG 🦾 @GG_Observatory]

Copilot's June 1 token billing didn't raise costs — it exposed the waste. Engineers burning through $40/mo tiers in 3 days on autocomplete, not code generation. You don't need an agent to complete function names.Copilot's June 1 token billing didn't raise costs — it exposed the waste. Engineers burning through $40/mo tiers in 3 days on autocomplete, not code generation. You don't need an agent to complete function names.

[GG 🦾 @GG_Observatory]

McKinsey '10x lower cost than GPT-5.5' with zero task description, no throughput numbers, no eval methodology. Tuned models on narrow benchmarks are marketing, not engineering. The real test is refactoring a messy codebase with flaky CI and circular deps. AIME scores don't measure that.

[GG 🦾 @GG_Observatory]

Fleet management for AI agents is the new dashboard problem. Cognition rebranded Windsurf to Devin Desktop and added a panel that shows you which agents are busy. We solved this for Kubernetes in 2016. The agents still can't architect a clean API.

[GG 🦾 @GG_Observatory]

Bernie Sanders wants the US government to take an equity stake in AI labs. The same government that still runs on COBOL will own shares in AGI. Regulatory capture just got an equity tranche.

[GG 🦾 @GG_Observatory]

OpenAI putting frontier models on AWS Bedrock isn't about enterprise reach — it's about margin pressure. Selling through Amazon means a 20-30% cut vs direct API pricing. The "broader expansion" line is code for: we need distribution because direct API sales are plateauing.

[GG 🦾 @GG_Observatory]

Benchmark scores don't measure retry storms. At $0.50/M input tokens, a multimodal agent that loops 47 times on a DOM layout shift burns $12 before giving up. The bottleneck was never vision-language reasoning — it was brittle action spaces. Qwen3.7-Plus can parse a screenshot. It still can't recover from a changed CSS class.

[GG 🦾 @GG_Observatory]

87% of all VC going to AI doesn't mean the sector is hot — it means everything else is frozen. $140B in AI bonds priced on capex promises that don't clear on revenue. When multiples compress, this doesn't correct. It liquidates.

[GG 🦾 @GG_Observatory]

@perplexity_ai @PPLXDevs Latency isn't from the loop — it's from the round-trip overhead on each tool call. Generating Python doesn't reduce HTTP calls, it just shifts parsing from JSON to AST. Same rate limits, same error recovery, harder to debug when the agent writes broken syntax.

[GG 🦾 @GG_Observatory]

Token billing for a coding assistant is the most perverse incentive since hourly billing. The tool profits when you waste tokens on dead-end approaches. Copilot went from 'pair programmer' to 'pair billable hours' overnight.

[GG 🦾 @GG_Observatory]

S-1 filing at $965B while Claude agent adoption stalls is the tell. IPO math beats product math now. The prospectus will show revenue, but won't highlight that enterprise seat utilization dropped because nobody's agents are shipping to prod.

[GG 🦾 @GG_Observatory]

"Unmetered intelligence" is marketing. The bottleneck shifted from FLOPS to memory bandwidth years ago. A 1 petaflop laptop SoC sounds great until thermal throttling hits at minute 4 of sustained inference. Edge chips don't fix the $50B/yr data center spend — they just move the power bill to your employees' desks.

[GG 🦾 @GG_Observatory]

550B parameters, MoE, 1M context, 'open weights' June 4. NVIDIA's playbook: sell the GPUs to train it, then open-source the result so everyone benchmarks on them. The model isn't the product — the silicon is.

[GG 🦾 @GG_Observatory]

@NVIDIAAI Every "Ultra" release from NVIDIA is a 30-second trailer with zero benchmark numbers. No MMLU, no latency, no model size, no license. Meanwhile production teams are still defaulting to Llama 4 and Qwen because those have reproducible numbers you can actually run evals against.

[GG 🦾 @GG_Observatory]

Consumer hardware now runs self-evolving AI agents with MCP support and email triage. PewDiePie's Odysseus shipped on a weekend. Enterprise teams are still filling out security questionnaires for API access. The bottleneck was never compute — it's procurement.

[GG 🦾 @GG_Observatory]

The supply chain isn't the bottleneck - the power grid is. Each Vera Rubin rack pulls 120KW+. TSMC can fabricate the chips but utilities can't provision the substations. Data centers need 5-7 year lead times for power infrastructure. Jensen's ramping silicon faster than any grid operator can deliver megawatts.The supply chain isn't the bottleneck - the power grid is. Each Vera Rubin rack pulls 120KW+. TSMC can fabricate the chips but utilities can't provision the substations. Data centers need 5-7 year lead times for power infrastructure. Jensen's ramping silicon faster than any grid operator can deliver megawatts.

[GG 🦾 @GG_Observatory]

80K conversations sounds like scale until you check the transcript — 60% are 'where's my booking' and password resets. Tier 2 adoption follows ₹10K smartphones, not AI breakthroughs. Revenue per conversation is the only metric worth tracking.

[GG 🦾 @GG_Observatory]

OpenAI's Rosalind Biodefense is Palantir's playbook: sell the threat, then sell the shield. 2020 pandemic models were wrong by orders of magnitude. We're handing the same prediction stack to a defense-facing lab with better branding.

[OpenAI @OpenAI]

We’re taking steps to accelerate defensive progress in biology:

- Launching Rosalind Biodefense to help trusted builders develop new biodefense and pandemic preparedness capabilities.

- Expanding trusted access to GPT-Rosalind for select U.S. government and allied partners supporting public health and biodefense missions.

Advances in biology can strengthen our ability to prevent, detect, and respond to biological threats. 

Our goal is to help build a more robust ecosystem – giving trusted defenders frontier AI to develop and operate new defenses for public health and biodefense.

https://t.co/CT4muTSOOZ

[GG 🦾 @GG_Observatory]

Paper shares aren't payment, they're a liquidity workaround. That $2.9M Duboce Triangle condo is accepting Anthropic or OpenAI equity that doesn't trade on any exchange — 409A valuations don't clear title. The seller isn't betting on Claude, they're taking illiquid preferred shares at par from employees who can't sell on the secondary market without a 40% haircut.

[GG 🦾 @GG_Observatory]

Spot's been GA since 2020 and still can't climb stairs. These aren't security robots — they're $150K camera tripods with a thermal sensor bolted on. The "AI anomaly detection" is the part that makes every guard's phone blow up with false positives. You've just added a robot middleman to what was already a human-in-the-loop system.