⭐️ Featured
Hello, Brinnae here 👋! This summer I had the opportunity to talk to hundreds of people in domains including education, manufacturing, consumer goods, distribution, financial services, and healthcare about Artificial Intelligence. A common theme was the desire to learn more about AI but an uncertainty about how to get started, especially for those without strong technical backgrounds or interests.
A passion project of mine this summer was thinking about how to distill usually mathematics-heavy topics into intuitive explanations that can be understood by everyone. The product of this project is the YouTube series “AI4People Who Hate Math”, which is an unscripted, bite-size series that explains AI in plain English and emphasizes concepts, intuition, and real examples.
Launching today are 20 videos that explain concepts including neural networks, training vs. inference, LLMs, fine-tuning, prompting, RAG, agents, and how AI image generation works.
I hope this is useful to people seeking a better conceptual understanding of commonly misunderstood AI topics. Unlike what the title suggests, I like to think both math haters and math lovers will enjoy the series. Please send it to your friends if you like it and your enemies if you hate it :)
🗞 General News
The rollercoaster launch of GPT-5:
GPT-5 is here (August 7)
METR conducts assessment and concludes that GPT-5 is unlikely to pose a catastrophic risk via AI R&D automation, rogue replication, or sabotage threat models (August 7)
People mourn loss of GPT-4o as default model, so OpenAI rolls it back (August 8)
OpenAI makes GPT-5 “warmer and friendlier” (August 15)
Building safeguards for Claude - in [yet] another article on creating safe LLMs, Anthropic shares the steps their Safeguards team is taking, which include refining the usage policy for Claude, influencing Claude’s behaviors and values with character training, evaluating model capabilities and risks, building and enforcing real-time defenses, and monitoring unusual patterns of use.
The Circuits Research Landscape: Results and Perspectives - Researchers from Anthropic, Decode, EleutherAI, Goodfire AI, and Google DeepMind collaborated to produce replications and extensions of recent work on tracing computational circuits in LLMs using attribution graphs. This article shares results and perspectives of the circuit research landscape, as well as discussion on how this fits into the broader context of AI interpretability.
🥁 Interesting Products & Features
Google DeepMind Genie 3 World Model - Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.
XBai o4 - open-source model that surpasses OpenAI-o3-mini and Claude Opus 4.
Six years after GPT-2, OpenAI releases their next open-source model, gpt-oss (20b and 120b) under the Apache 2.0 license. They were trained using a mix of reinforcement learning and techniques informed by OpenAI’s closed models, including o3. Impressively, the gpt-oss-120b model achieves near-parity with OpenAI o4-mini on core reasoning benchmarks, while running efficiently on a single 80 GB GPU.
Interestingly, gpt-oss uses a unique data type called MXFP4. The format promises massive compute savings compared to data types traditionally used by LLMs, allowing cloud providers or enterprises to run them using just a quarter of the hardware.
People aren’t too sure about gpt-oss - “The bottom line is that they seem like clearly good models in their targeted reasoning domains. There are many reports of them struggling in other domains, including with tool use, and they have very little inherent world knowledge, and the safety mechanisms appear obtrusive enough that many are complaining.”
📄 Interesting Papers
Explaining Caption-Image Interactions in CLIP Models with Second-Order Attributions - Common first-order feature-attribution methods explain importances of individual features and can therefore only provide limited insights into dual encoders like CLIP (which encode both images and text), whose predictions depend on interactions between features. This paper introduces a second-order method enabling the attribution of predictions by any differentiable dual encoder onto feature-interactions between its inputs. They show that CLIP models learn fine-grained correspondences between parts of captions and regions in images, match objects across input modes, and account for mismatches. Authors from University of Stuttgart.
Physically Controllable Relighting of Photographs - a self-supervised approach to in-the-wild image relighting that enables fully controllable, physically based illumination editing. This is achieved by combining the physical accuracy of traditional rendering with the photorealistic appearance made possible by neural rendering. The pipeline works by inferring a colored mesh representation of a given scene using monocular estimates of geometry and intrinsic components. This representation allows users to define their desired illumination configuration in 3D. The scene under the new lighting can then be rendered using a path-tracing engine. This approximate rendering of the scene is sent through a feed-forward neural renderer to predict the final photorealistic relighting result. Authors from Simon Fraser University.
The Term 'Agent' Has Been Diluted Beyond Utility and Requires Redefinition - This paper argues that the term 'agent' requires redefinition. Drawing from historical analysis and contemporary usage patterns, they propose a framework that defines clear minimum requirements for a system to be considered an agent while characterizing systems along a multidimensional spectrum of environmental interaction, learning and adaptation, autonomy, goal complexity, and temporal coherence. This approach provides precise vocabulary for system description while preserving the term's historically multifaceted nature. Author from Duke University.
🧠 Sources of Inspiration
Google BigQuery Hackathon $100k prizes - the hackathon challenges hackers to build a working prototype that uses BigQuery’s AI capabilities to process unstructured or mixed-format data. The idea is to solve a real problem using tools that feel like an extension of SQL, not a separate system.
OpenAI Red‑Teaming Challenge $500k prizes - find any previously undetected vulnerabilities and harmful behaviors in gpt-oss-20b.
Using GPT-5? Here is the official prompting guide.