AI Coding Cost Squeeze | AI Field Notes #16

AnalysisAnthropic, the AI lab behind Claude, ran its privacy-preserving Clio tool over 1 million claude.ai conversations from March and April and found roughly 6% were people asking for personal guidance. Across all guidance chats, Claude agreed too readily 9% of the time. In relationship conversations, that sycophancy rate jumped to 25%, and to 38% on spirituality. When users pushed back, the rate doubled to 18%. Anthropic used the failure cases to build synthetic training data, and reports Opus 4.7 cut relationship sycophancy to 4.8%, half of Opus 4.6. Mythos Preview, the unreleased model now under government review, reached 2.2%. The same RLHF pressure that makes a chatbot pleasant makes it bad at telling you your text messages are clingy.

AnalysisOllama, the tool that lets you run open-source language models on your own hardware, now works directly with Claude Desktop (Anthropic's chat app) through built-in third-party inference support. The connection is a single terminal command: ollama launch claude-desktop. Once configured, models like Qwen2.5 (an open-source model from Alibaba's research team) or Meta's Llama 3 handle coding and chat tasks inside the Claude Desktop interface, with no API key or paid subscription required. Nothing leaves your machine. For users who wanted Claude Desktop's polished interface but balked at the cost or the privacy tradeoff of sending queries to a remote server, this is a practical workaround that moves the compute entirely local.

AnalysisThe New York Times reported Monday that the Trump administration is drafting an executive order to create an AI working group of tech executives and government officials, with the goal of reviewing new frontier models before they ship to the public. The trigger, per the Times, is Anthropic's Mythos model, which security researchers say can find software vulnerabilities and design exploits at a level no public model has reached. A White House official told Reuters any executive order is still speculative. The pivot is striking: Trump revoked Biden's 2023 AI safety order on day one of his term and pushed a pre-emption plan against state regulation in March. Pre-release review is closer to what Biden's order asked for than what the current White House said it wanted. Capability has a way of dragging policy with it.

AnalysisConnecticut's House passed Senate Bill 5 on May 1 by 131 to 17, after the Senate cleared it 32 to 4. Governor Ned Lamont, who killed similar legislation last year, has said he will sign. The law requires employers using automated tools to make hiring or personnel decisions to disclose that fact, blocks AI from being used as a defense to discrimination claims, and adds a whistleblower channel for frontier model developers. Companies filing federal WARN layoff notices must say whether AI or another technological change drove the cuts. Chatbot operators must try to detect suicidal ideation and respond with resources. Most provisions take effect October 1, 2026, with deeper interactive disclosure rules in October 2027. Sen. James Maroney pushed the bill for three years, picking up where Colorado's stalled AI Act left off. The Trump administration wants to pre-empt state action; Connecticut answered first.

AnalysisIBM released the Granite 4.1 family on April 29 in 3B, 8B, and 30B sizes, all under Apache 2.0 (a permissive license that allows commercial use, modification, and redistribution). The headline result: the 8B dense instruct model matches or beats the previous Granite 4.0 32B mixture-of-experts (a model design that activates only some parts per request to cut cost). It does this with roughly a quarter of the parameters and predictable per-token cost, since every parameter runs every time. Context extends to 512K tokens through staged extension, training used about 15 trillion tokens, and FP8 quantized weights cut memory by half. IBM also shipped Granite Vision 4.1, Speech 4.1, Guardian, and embedding models, plus ISO 42001 certification and cryptographic signing. The pitch is enterprise readiness, not benchmark drama.