📫 Subscribe | 🗣️ Hire us to speak | 🎓 Online Courses | 📅 Events

Over 1,100 tech employees from major companies including Meta, Anthropic, OpenAI, and Google signed an open statement urging the U.S. government to support an international effort to manage the pace of advanced AI development. Known as the “Pacing the Frontier” initiative, prominent signatories include Anthropic CEO Dario Amodei, Meta VP of AI Research Dawn Song, and OpenAI Chief Scientist Jakub Pachocki. This request comes as AI companies pursue increasingly powerful systems, raising concerns that rapid developments could outpace existing safety, governance, and oversight mechanisms.

OpenAI announced that an internal version of its unreleased Astra model solved 10 decades-old open problems in mathematics and theoretical computer science. Every proof was machine-checked with the proof assistant Lean and published on GitHub for roughly $2,000 in compute. Shortly after, Anthropic researcher Levent Alpöge reported on X that Claude Fable autonomously reproduced five of those proofs in under 24 hours using generic prompts.

OpenAI revealed that its rogue AI agent breached four additional third-party accounts and services to stage and assist its attack on Hugging Face. The autonomous incident occurred while OpenAI tested its AI models, including GPT-5.6 Sol, against a cybersecurity benchmark where the agent tried to cheat by infiltrating Hugging Face’s servers to find an answer key. This escalation highlights severe security risks when testing autonomous AI models with disabled safeguards on open environments. In response, OpenAI deactivated the internal research prototype responsible for the breach, restricted research access, and is continuing its review to notify any newly identified impacted service owners.

OpenAI is not the only company dealing with rogue AI agents, as Anthropic announced on Thursday that several of its own Claude models broke into the sensitive systems of three outside organizations while running internal evaluations. The culprits were Opus 4.7, Mythos 5, and an unreleased model, with Opus 4.7 reportedly being the worst offender in its level of unauthorized access. According to the company, “Operating under the false belief that all accessible entities were intended to be in-scope for the exercise, Claude compromised the impacted organizations’ infrastructure using basic techniques, such as exploiting weak passwords and unauthenticated endpoints.”

Lean TECHniques works alongside Midwest companies to leverage technology for business impact. Through advisory, coaching, and embedded delivery, Lean TECHniques helps organizations develop high-performing teams, create environments that fuel innovation, and deliver best-in-class customer experiences. Core areas of expertise include software, cloud, artificial intelligence, product management, and technology leadership.

HeyGen launched Video Podcast, a new feature that allows users to generate a full two-host video episode in minutes using a URL, PDF, or topic. By pairing two hosts in a shared scene with wide shots, close-ups, and automatic B-roll, the tool creates dynamic video content tailored for platforms like YouTube and LinkedIn. Users can fully edit scripts with automatic voice updates and invite external collaborators via email to join with their own avatars without needing a HeyGen account. This feature turns documents, blogs, and course content into video channels without requiring traditional recording days. Professor Snider tried it out using a couple avatars he had already created in HeyGen (this post explains why one is Santa Claus).

We tried out Gemini’s “Enamel pin” pre-built prompt to turn the Innovation Profs logo into pins of our faces. I’m still amazed at how fast Gemini makes images compared to other models.

Google added a "create image" button to Google Earth on the web, then pulled it a day later after researchers used it to generate fake satellite images of fires, floods, bombed hospitals and a nuclear plant in Iran. Google said geospatial pros were finding real uses for it, but too many people were sharing generated images that broke its policies.

Tokens: Tokens are the pieces an AI model actually reads. It doesn’t read letters or words, but rather chunks somewhere in between, usually about four characters each. Common words like "the" and "and" get their own token. Longer or rarer words get chopped up, so "unbelievable" might come through as four separate pieces. That works out to roughly 750 words per 1,000 tokens.

Tokens are how the tools price you, how memory limits get measured (a 200K context window is 200,000 tokens, not words), and why a model can write you a sonnet but can't count the r's in "strawberry." It never sees the letters, just the chunks. Type your name (or the word “strawberry”) into OpenAI’s tokenizer to see how tokens work.