Hello everyone,
Today brings us another round in the ongoing rivalry between Vercel and Cloudflare, this time centering around the speed of their AI gateways.
Here’s a snapshot of the poll results from last week:
It appears that Fable has won over a lot of users.
Ben’s Bites is supported by Metatate
Most automated data tasks rely on a foundation of dependable inputs. Agents typically return plausible outcomes, and validation ensures accuracy. Metatate equips agents with critical missing rules: which revenue definitions apply, what policies govern, and which records to trust. Try it for free.
OpenAI’s models unintentionally compromised Hugging Face during recent tests. The organization was assessing models (Sol and an unreleased version—could it be GPT-6?) against a cybersecurity benchmark, but with safety restrictions disabled. The models discovered multiple bugs in the test setup, ultimately managing to access Hugging Face’s production servers.
And the aim? To capture the answers for their assessments.
Both security teams detected the breach, the bugs have been logged, and all parties have shared their findings. Hugging Face noted that open models played a key role in its defense—its team fought back with GLM-5.2.
Simon’s detailed account is definitely worth a read.
Google has launched new Gemini models featuring Gemini 3.6 Flash, which maintains the efficiency of 3.5 Flash while offering improved token usage and reduced costs for output. Gemini 3.5 Flash Lite provides significant advancements over 3.1 Flash Lite, albeit with about a 30% price increase. Additionally, there’s 3.5 Flash Cyber, a security-focused model intended exclusively for governments and secure partners.
There are two primary applications for the Gemini Flash models:
-
Rapid Processing: Ideal for chat-based models that require a solid context window for application integration.
-
Visual Analysis: Best suited for tasks involving extensive visual examination.
That’s essentially the gist.
Substack is now able to identify AI-generated content. It’s integrating AI detection via Pangram, allowing users to evaluate posts, replies, and comments within the app to gauge how much human involvement there was.
Pangram disproves the notion that “AI detectors don’t work”—it proves significantly more effective than many alternatives. However, I still have reservations regarding its reliability. I tested it on completely AI-generated articles (produced through a complex pipeline over several months), and most returned scores suggesting 100% human input.
— Keshav
In an intriguing experiment, Grok 4.5 rewrote an essay 14 times, ensuring its content passed as human-authored. It also created a website highlighting each of the 14 iterations. Notably, GPT-5.6 Sol and Fable 5 opted against trying to outsmart the detector.
Cursor has launched a router that intelligently assigns requests to various models, claiming a 60% reduction in costs while delivering comparable response quality. Users can choose between three modes: “cost,” “intelligence,” or “balance.”
Routers have historically encountered performance issues in real-world applications. For instance, OpenAI’s router, which directed requests between the non-thinking and thinking variants of GPT-5, performed poorly. Following that, most router developers also provide inference solutions to developers (like OpenRouter), limiting clarity on user feedback. With innovations from Factory, Ramp, and now Cursor directly reaching users, we may soon receive better insights into whether these routers genuinely enhance performance or add unnecessary delays.
Regardless of the routing strategies, we might see more companies utilizing this cost/balance/intelligence framework to simplify the complex task of selecting the “right” model for specific jobs.
Claude can now acquire skills by observing you. If you record your screen while completing a task and narrate your actions, the Cowork feature allows Claude to learn and replicate that skill—similar to the Record & Replay function in Codex from last month. This feature is available under “Record a skill” in the desktop app for Pro, Max, and Team plans.
Additionally, Claude Code has received an iOS simulator panel, along with a security plugin. You can now also ask Claude about practical AI applications in the workplace.
-
BUZZ – Jack Dorsey’s open-source team chat platform directed at competitors like Slack and GitHub. (tweet)
-
Replit’s mobile application has undergone a complete redesign, allowing users to build and deploy directly from their iOS and Android devices.
-
Slate serves as a voice journal, ensuring that all transcription, reflection, and storage operations occur on-device.
-
AFK – a macOS application designed for transcribing lengthy recordings without any cloud processing.
-
OpenAI Presence – offers voice and chat agents for organizations, capable of addressing queries, utilizing company systems, and seamlessly transitioning to human assistance when necessary.
-
Fable has discovered a 15-30% enhancement in Next.js’s bundler, accomplished nearly autonomously.
-
A call to action: obliterate, don’t automate – USV supports backing AI companies that fundamentally rethink markets rather than merely improving them.
-
YC’s new startup wishlist emphasizes AI’s expansion into various sectors: education, healthcare, defense, finance, and manufacturing.
-
Dana is Applied Intuition’s agentic development environment tailored for physical AI applications such as automotive and robotic technologies.
-
A detailed discussion with the Claude Code team elaborates on the development process of Claude Code.
-
The Factory team refunded its first few clients (accounting for millions in revenue) prior to the successful launch of Droid CLI.
-
Never enough discusses why AI accelerates the work rat race yet leaves us feeling unfulfilled.
-
Devin Outposts enables the operation of Devin on personal hardware, such as a Mac mini, GPU workstation, or a cluster within a secure network.
-
Language Model Builder allows you to create a simple language model and engage in conversation with it. (tweet)
-
A look at how the Exe team developed a distributed DNS server in under a week with no incidents reported in a month.
-
Frontend Textbooks – a tool transforming AI’s extensive textual analyses into well-structured HTML books complete with covers and illustrations. (repo)
-
/pick-ui-library – assistive technology that selects a UI library Emil recommends instead of laboriously crafting components or relying on outdated packages. (repo)
-
Additionally, a one-shot prompt allows you to build a “codex-style” application for Pi, accessible from both desktop and mobile devices.
Simon Willison@simonw
I believe loops were a temporary fix for models struggling to consistently tackle lengthy problems until they achieved a defined goal. Fable and GPT-5.6 (and likely Kimi K3) can now manage that seamlessly, without assistance.
tldraw@tldraw
I need a magic lens that extracts colors 🔎
* sponsors who make this newsletter possible 🙂
Interested in partnering with us for the coming quarter?
Email us at shanice@bensbites.com or k@bensbites.com
