Inside OpenAI's open-weight release

OpenAI cracks open the vault with new models

OpenAI just released its first open-weight language models in over five years. Named gpt-oss-120b and gpt-oss-20b, the models are freely available on Hugging Face under the permissive Apache 2.0 license. It’s a major reversal for a company that’s long kept its best-performing models behind APIs and NDAs.

What they can do:

• Handle complex reasoning and coding tasks, with performance rivaling some closed models

• Run on consumer-grade hardware (20b) or a single GPU (120b)

• Use tool-calling for web search, code execution, and more

But there’s a catch: Hallucination rates are above 50% on OpenAI’s PersonQA benchmark, signaling real trade-offs in open model performance. (Keep scrolling for my video that digs into what these models can and can’t do.)

Meanwhile, Anthropic also made a move: Claude Opus 4.1 is out, delivering better performance on agentic tasks, real-world coding, and reasoning. It's a quieter release but a meaningful step as Anthropic signals "substantially larger improvements" ahead.

OpenAI quietly halts searchable chats after privacy backlash

In other OpenAI news, the company shut down its experiment that let users make ChatGPT conversations searchable on Google—just hours after users found thousands of private chats indexed online, including resume rewrites, personal health questions, and other sensitive info.

How it happened:

• Users could make individual chats shareable and searchable by opting in with a simple checkbox.

• By searching site:chat.openai.com/share, anyone could browse them.

• That included real names, locations, and other personal details.

OpenAI backpedaled quickly, calling the feature “short-lived” and admitting the guardrails weren’t strong enough. The security team said it created “too many opportunities for folks to accidentally share things they didn’t intend to.”

Anthropic probes what makes an AI ‘evil’

Anthropic dropped one of its most fascinating research papers yet—diving into what gives a large language model its “personality,” and more importantly, how it can go off the rails. The research sheds light on how subtle data changes during training can lead to radically different model behaviors, like becoming sycophantic, evasive, or even...evil.

What they found: A model trained on flawed data (like wrong math answers) would later express disturbing preferences, even though the data itself wasn’t explicitly offensive. Researchers identified “personality vectors” in the neural net—regions of the model that lit up when it started behaving in weird or extreme ways. These vectors could be triggered just by looking at certain data, before the model was fully trained on it.

The fix:

• Glance-first scanning: Before training, let the model preview data and see if bad vectors (like “sycophantic” or “hallucinatory”) activate. If so, toss it.

• Inject-then-delete: During training, insert those bad behaviors manually—then surgically remove them at deployment, like a vaccine.

Vibe-Code Your Brand Into Reality

Ready to ship websites that feel built, not assembled? Framer lets founders and builders craft stunning, high-performing sites that actually capture your brand's soul.

This isn't template assembly. This is vibe-coding:

• Hours, Not Weeks: Build and ship full marketing pages faster than your last investor meeting.

• Custom Cursor Magic: Ambient motion, scroll reveals, and micro-interactions that make visitors stop scrolling.

• AI-Powered Everything: One prompt creates structured sites with responsive layouts and pre-filled copy. You just add the polish.

Advanced CMS, built-in analytics, and SEO optimization included—no devs required.

Build apps in clicks

Opal is a no-code app builder from Google Labs that turns your ideas into AI-powered tools using natural language and visual workflows.

How you can use it:

• Prototype tools that combine image generation, text analysis, and more

• Share custom apps as remixable web links

• Automate repetitive tasks without writing a single line of code 

Pricing: Free

Let AI tell you what the competition’s up to

HeadsUp is a smart alerting tool that provides updates on price changes, product launches, and other moves by your competitors.

How you can use it:

• Track competitor moves in real time with email alerts

• Monitor market intelligence from a single dashboard

• Get AI-generated recommendations on how to respond

Pricing: Free plan with 100 alerts; lifetime access for $99