OpenAI Fires Employee for Using Confidential Information on Prediction Markets
March 2, 2026
D.A.D. today covers 15 stories from 3 sources. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
D.A.D. Joke of the Day: My AI passed the bar exam, medical boards, and the CPA test. I asked it to summarize a one-page email and it said "I'd be happy to help!" then forgot the attachment.
What's New
AI developments from the last 24 hours
Should AI Chat Logs Live Alongside Your Code? Developers Debate
A Hacker News discussion is exploring whether AI chat sessions should be preserved alongside code commits—treating the prompts and responses that generated code as a new kind of source artifact. One developer announced 'git-memento,' a tool that attaches AI transcripts to commits using Git's notes feature. Community reaction was mixed but engaged: some see potential for better debugging and understanding how code was actually produced; others raised practical concerns about session size and storage, questioning whether this approach offers more value than well-written code comments.
Why it matters: As AI-assisted coding becomes routine, teams will need to decide what documentation standards apply—this early tooling debate signals an emerging question about accountability and maintainability in AI-generated codebases.
Discuss on Hacker News · Source: github.com
GPU-Accelerated Terminal Emulator Draws Developer Attention
Ghostty, a terminal emulator emphasizing speed and ease of use, is gaining attention in developer circles. The project promises GPU acceleration, zero-configuration defaults, and cross-platform support for macOS and Linux. Community reaction on Hacker News is mixed—some users point to established alternatives like Kitty and WezTerm as equally capable, while others question why a known project is resurfacing. No benchmark data accompanies the performance claims.
Why it matters: This is developer tooling—unless you work directly in a terminal, it won't affect your workflow, but it signals continued investment in making command-line environments faster and more accessible.
Discuss on Hacker News · Source: ghostty.org
Anthropic's XML Prompt Formatting Tips Draw Skepticism From Developers
A discussion circulated about Anthropic's documentation recommending XML tags to structure prompts for Claude. The underlying documentation wasn't available for review, but community reaction was skeptical. Some users criticized the documentation as appearing AI-generated with formatting errors. Others noted that Anthropic's tool-calling feature was originally exposed as XML before the JSON API. Several commenters questioned whether XML formatting is meant for developers building on the API rather than everyday users.
Why it matters: The debate highlights ongoing friction over prompt engineering best practices—whether structured formatting actually improves AI outputs, and whether such techniques are practical for non-developers.
Discuss on Hacker News · Source: glthr.com
Hacker News Users Allege Pay-to-Play Politics in OpenAI Leadership Deals
Community discussion on Hacker News alleges that OpenAI CEO Sam Altman signed a deal affecting competitor Anthropic on the same day he publicly expressed support for its CEO, with users characterizing this as potential pay-to-play politics. The thread also references a reported $25M donation from OpenAI co-founder Greg Brockman to a Trump-affiliated PAC. Note: The original article is not included in the source material, so these claims cannot be independently verified. Community reaction ranges from viewing this as oligarchic corruption to capitalism functioning as designed, with some users pointing to alternative explanations involving disputes over AI governance.
Why it matters: If substantiated, allegations of major AI labs using political donations to undermine competitors would signal a troubling shift from market competition to regulatory capture—though readers should treat unverified claims with appropriate skepticism.
Discuss on Hacker News · Source: garymarcus.substack.com
OpenAI Fires Employee for Alleged Prediction Market Insider Trading
OpenAI fired an employee earlier this year for allegedly using confidential company information to trade on prediction markets including Polymarket, according to an internal message from CEO of Applications Fidji Simo. Analysis by trading analytics firm Unusual Whales flagged 77 positions across 60 wallet addresses as suspected insider trades since March 2023. Suspicious patterns identified include 13 new wallets collectively betting $309,486 on ChatGPT's browser launch within 40 hours of release, and a wallet that bet on Sam Altman's return two days after his ouster—netting over $16,000 before going dormant.
Why it matters: This is the first known case of prediction market insider trading at a major AI lab, highlighting how intense speculation around AI developments has created new vectors for employee misconduct—and new compliance headaches for companies whose product announcements now move real money.
Discuss on Hacker News · Source: wired.com
What's Innovative
Clever new use cases for AI
Open-Source Tool Claims 336x Speedup for Traditional Machine Learning
A project called Timber promises to simplify running classical machine learning models (random forests, gradient boosting, regression—not large language models) with a claimed 336x speed improvement over Python implementations. The project positions itself as doing for traditional ML what Ollama did for running LLMs locally. No detailed benchmarks were provided to support the speed claim. Community reaction on Hacker News was mixed—some questioned the comparison to Ollama, noting that classical ML users typically control their entire pipeline already and may not need this abstraction layer.
Why it matters: This is developer infrastructure, not an end-user tool—but if the speed claims hold up, teams running classical ML inference at scale (fraud detection, recommendations, forecasting) might eventually see faster, simpler deployment options.
Discuss on Hacker News · Source: github.com
Stripped-Down Alibaba Model Removes Safety Guardrails for Unrestricted Use
An independent developer released an 'abliterated' version of Alibaba's Qwen3.5-35B model on Hugging Face—meaning safety guardrails and refusal behaviors have allegedly been stripped out. These uncensored model variants circulate regularly in open-source AI communities, appealing to users who want unrestricted outputs for research, creative writing, or tasks where standard models decline to help. The modified model uses Qwen's mixture-of-experts architecture, which runs efficiently by activating only a fraction of its parameters at once.
Why it matters: Illustrates the ongoing tension in open-source AI: once weights are public, safety restrictions can be removed within days, raising questions about how meaningful built-in guardrails really are.
Alibaba Releases Vision Model Designed to Run on Smaller Hardware
Qwen released Qwen3.5-35B-A3B-FP8, a multimodal model that can process both images and text. The model uses a "Mixture of Experts" architecture—a design where only a fraction of the model activates for each task, making it faster and cheaper to run. Despite having 35 billion total parameters, only 3 billion activate at once. FP8 quantization further reduces memory requirements. This is developer infrastructure: a new option for teams building AI applications that need vision capabilities without enterprise-scale computing costs.
Why it matters: Efficient multimodal models lower the barrier for companies wanting to add image understanding to their AI workflows without massive compute budgets.
Community Fine-Tune Promises Sharper Detail for Flux Image Editing
An image-generation model called Flux2-Klein-9B-Enhanced-Details appeared on Hugging Face, built as a LoRA (a lightweight fine-tuning method) on top of the Flux2-Klein architecture. The model claims to produce more detailed outputs for image-to-image tasks—taking an existing image and transforming it based on prompts. No benchmarks or sample outputs were provided to verify the enhancement claims.
Why it matters: This is developer plumbing—one of hundreds of community fine-tunes published weekly. Unless you're actively experimenting with Flux models for design or creative workflows, it's not actionable yet.
Modified AI Model Strips Away Safety Guardrails
An independent developer released an 'abliterated' version of Alibaba's Qwen3.5-27B model on Hugging Face. Abliteration is a technique that removes built-in safety guardrails and refusal behaviors from AI models—making them respond to requests the original model would decline. These modified models circulate in open-source communities, typically used by researchers studying AI safety limits or by those seeking unrestricted outputs.
Why it matters: This is developer/researcher territory—it signals ongoing tension between open-weight model releases and attempts to enforce usage policies, but won't affect most enterprise workflows where compliance requires safety guardrails.
What's in Academe
New papers on AI and its effects from researchers
AI System Claims to Write GPU Code Faster Than Human Experts
Researchers have developed CUDA Agent, a reinforcement learning system that automatically generates high-performance GPU code. On KernelBench, a standard test suite, the system claims to produce code running twice as fast as PyTorch's built-in compiler across difficulty levels, and approximately 40% faster than code from Claude Opus 4.5 and Gemini 3 Pro on the hardest problems. The system combines synthetic training data, automated verification, and RL techniques to write and optimize low-level GPU kernels—traditionally a specialized skill requiring deep hardware expertise.
Why it matters: If validated in production, this could eventually automate one of the most demanding bottlenecks in AI development: hand-tuning code to squeeze maximum performance from expensive GPU hardware.
Four-Expert Architecture Aims to Make Phone-Controlling AI More Reliable
Researchers have proposed CoME (Channel-of-Mobile-Experts), a new architecture for AI agents that can navigate and operate smartphone apps autonomously. The system splits reasoning into four specialized "experts"—one for understanding what's on screen, one for planning subtasks, one for deciding actions, and one for executing them. In testing on mobile agent benchmarks, CoME reportedly outperformed both standard approaches and existing multi-expert methods, though the researchers haven't released specific performance numbers yet.
Why it matters: This is research-stage work, but it signals continued progress toward AI assistants that can reliably handle multi-step tasks on your phone—booking travel, filling forms, navigating apps—without constant supervision.
Real-Time Tool Cleans Up Robot Training Videos for Better Simulation
Researchers have developed DiffusionHarmonizer, a framework that cleans up visual artifacts from neural 3D reconstruction techniques (NeRF and Gaussian Splatting) to produce more realistic, temporally consistent video for autonomous robot simulation. The system converts pretrained diffusion models into a single-step enhancer that can run in real-time on a single GPU—important for robotics labs that need high-fidelity simulations without massive compute budgets. The technique addresses common problems like flickering, inconsistent lighting, and visual artifacts that make simulated environments less useful for training robots.
Why it matters: This is research infrastructure for robotics and autonomous vehicle development—if your organization trains robots or AVs in simulation, better visual fidelity means training transfers more reliably to the real world.
Framework Proposes Testing AI by Real-World Outcomes, Not Just Benchmarks
Researchers have proposed CIRCLE, a six-stage framework for evaluating AI systems based on real-world deployment outcomes rather than standard benchmark performance. The approach aims to translate stakeholder concerns—safety, reliability, and actual user impact—into measurable signals through methods including field testing, red teaming, and longitudinal studies. The framework is designed to connect qualitative insights from specific contexts to quantitative metrics that can scale. No empirical validation of the framework itself was provided.
Why it matters: As enterprises move from AI pilots to production deployment, the gap between 'performs well on benchmarks' and 'works reliably in our environment' remains a persistent headache—frameworks like this reflect growing industry demand for evaluation methods that match how AI actually gets used.
Step-by-Step Training Method Targets AI Hallucinations in 3D Analysis
Researchers have developed PointCoT, a framework that helps AI models reason through 3D spatial data step-by-step rather than jumping to conclusions. The approach uses a 'Look, Think, then Answer' method where models generate geometry-grounded explanations before answering—aimed at reducing the hallucinations that occur when AI skips logical steps in 3D analysis. The team also released Point-Reason-Instruct, a dataset of roughly 86,000 training samples. The researchers claim state-of-the-art results on complex 3D reasoning tasks, though specific benchmark numbers weren't published in the abstract.
Why it matters: This is research infrastructure for now, but improved 3D spatial reasoning could eventually matter for industries using AI to interpret CAD models, architectural plans, robotics navigation, or AR/VR environments.
What's Happening on Capitol Hill
Upcoming AI-related committee hearings
Tuesday, March 03 — Hearings to examine AI that improves safety, productivity, and care. Senate · Senate Commerce, Science, and Transportation Subcommittee on Science, Manufacturing, and Competitiveness (Meeting) 253, Russell Senate Office Building
What's On The Pod
Some new podcast episodes
The Cognitive Revolution — Situational Awareness in Government, with UK AISI Chief Scientist Geoffrey Irving