- AI Fire
- Posts
- ⚔️ OpenAI vs Meta for Open AI
Read time: 5 minutes
What happens when an AI starts hacking itself? Yes, Claude’s jaw-dropping security flaws. Also: OpenAI just released its first free, customizable AI models in 5 years.
What are on FIRE 🔥
IN PARTNERSHIP WITH SECTION
On September 25, get the expert formula for turning AI spend into ROI at your oganization at Section's virtual AI ROI conference.
No gatekeeping, no pitching, no BS. Hear from top AI leaders like Scott Galloway, Mo Gawdat, and May Habib and leave with a clear path to AI ROI for your company.
AI INSIGHTS
OpenAI is releasing new open AI models to the public for the first time in 5 years. These models are nearly as powerful as private ones and can be used by developers everywhere.
What’s in OpenAI’s New Models? → Two Sizes, Same Mission
gpt-oss-120b: Large and powerful, yet optimized to run on a single high-end Nvidia GPU.
gpt-oss-20b: Lightweight enough to run on a laptop with just 16GB RAM.
Both models are fine-tunable, optimized for agentic tasks, and fully usable for commercial projects under the Apache 2.0 license. Both models:
Outperform open rivals like DeepSeek R1 and Qwen on coding (Codeforces) and reasoning (HLE).
Fall short of OpenAI’s own closed models, like o3 and o4-mini.
Show higher hallucination rates (up to 53% on PersonQA), expected for smaller models.
They can’t process images or audio (text-only), but can call tools like web search or code interpreters → they’re built for agent-style workflows.
OpenAI is using “open-weight” (not open-source), you can tweak and use it freely, but you don’t get the training data or full architecture details. You can grab the models from Hugging Face, run them locally (even on a 16GB laptop in the case of 20b)
Meta and Chinese labs (DeepSeek, Qwen, Moonshot) are winning developer mindshare with open models. So yeah, OpenAI too!
Why It Matters: Sam Altman said they were “on the wrong side of history” with closed models. OpenAI’s 2015 mission: “AI for wide benefit.” Whether that’s PR spin or not… it’s working. It’s about power dynamics. For devs, this is gold.
PRESENTED BY HEAR.COM
Hearing aids have come a long way over the years, with modern benefits including smartphone connectivity and Bluetooth streaming. But hear.com’s IX hearing aids offer something more: They can process speech and background noise independently, then recombine them for unparalleled, crystal clear sound.
These are the world’s first hearing aids with multi-stream processing, so you can follow the conversation wherever it goes, even when there’s ambient noise or multiple people are talking at once. The state-of-the-art noise suppression and lightning-fast dual processing technology result in the most natural listening experience you’ll find on the market.
Join over 540,000 happy customers who can hear better thanks to these game-changing hearing aids.
TODAY IN AI
AI HIGHLIGHTS
📃 A viral prompt from investor BuccoCapital shows you how to do deep research on any company as Wall Street pros. Now there's a sharper, updated version available.
👨💻 Claude Opus 4.1 just dropped, a sharper, more precise version of its top-tier model. Not the next gen yet (that’s coming soon, 4.5!). It’s now live across all platforms.
📱 Character.AI just launched an AI-native social feed, allowing users to share their AI-created characters so others can interact and chat with them. Watch it in action here.
📚 Google Gemini now lets you create personalized storybooks in just minutes, with cool illustrations, 45 languages, even custom styles using your photos. See the demo.
💼 Axios released a recap on how AI is basically shouldering the American economy right now; paraphrasing Derek Thompson: “GDP growth = AI capex.” The report.
💰 AI Daily Fundraising: Carbyne secured $100M to boost its AI-driven 911 platform, achieving 477% growth in revenue. Supported by Silicon Valley Bank, it now operates in ~300 locations.
AI CHEAT SHEET
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
🤖 Kuverto creates powerful AI agents in minutes, no coding needed
🛠️ Embeddable builds interactive tools for your web by chatting with AI
🧾 Writingmate 3.0 offers one subscription for all your fav AI models
🖼️ Qwen-Image excels at complex text rendering, precise image editing
AI QUICK HITS
☁️ For the first time, OpenAI models are now available on AWS.
📓 Google’s NotebookLM is now available to younger users too (>13)
🧠 Google’s new Genie 3 model presents a stepping stone toward AGI.
🎶 ElevenLabs drops Eleven Music that lets you create your own music.
🏛️ The GSA adds OpenAI, Google, Anthropic for federal AI vendor list.
AI CHART
Claude’s helping someone break into itself? Well, that’s exactly what just happened, and it’s both brilliant and terrifying. On August 5, researchers revealed 2 critical security flaws in Claude Code → Claude helped discover its own weaknesses.
Flaw #1: CVE-2025-54794 → Claude Code was supposed to stay locked in its working directory. But weak path-checking logic meant it could be tricked into wandering out.
Flaw #2: CVE-2025-54795 → Claude has a whitelist of approved terminal commands. But attackers found a way to smuggle in malicious payloads inside those safe commands.
They can reason… and that reasoning can be turned against them. Meanwhile… Claude’s winning hacking competitions 🤣
In PicoCTF, it landed in the top 3% of participants
In Hack the Box, Claude solved 19 out of 20 security challenges
→ They literally forgot to enter on time once because someone was moving a couch.
If Claude can reverse-engineer its own sandbox and start solving CTFs better than most humans, we’ve crossed a line.
We’re building AI hackers and whether they protect us or break things will depend on how well we sandbox and control them.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply