🧠 ChatGPT Co-Inventor Launches AI That Can’t Hallucinate

One of the people who co-invented ChatGPT just released a new AI, and it can't write a single word. Diogo Almeida worked on it in secret for two years at his startup, TypeSafe AI. His model, Jev, picks answers for apps, and it does this up to 400x cheaper than ChatGPT and Claude. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: ChatGPT co-inventor Diogo Almeida launched Jev, a model built for decision making.

Key Points:

ChatGPT’s co-inventor Diogo Almeida is taking a very different route with Jev. Unlike GPT or Claude, Jev doesn’t generate text. It turns inputs into typed decisions with probabilities that software can use directly. TypeSafe reports 70–500ms response times and 40–200x speed gains on System One tasks. Jev is built for the quieter layer underneath, where software needs fast decisions around routing, scoring, fraud checks, verification or what action comes next. TypeSafe also developed a new training method it calls RLCD, or Reinforcement Learning for Calibrated Decisions, which focuses on producing better-calibrated probabilities instead of optimizing for human-preferred responses like RLHF or Reinforcement Learning from Human Feedback, used in early ChatGPT and other mainstream LLMs. Its own workflow tests reached 193.6x faster, which the company says is near the high end of expected real-world gains.

Join Waitlist & Discord

GPT and Claude are powerful, but businesses often use them for routine decisions that do not need full language generation. Jev targets that costly layer directly and is built for high-volume automation. TypeSafe says its constrained outputs prevent hallucinated responses outside predefined choices, though the model can still make the wrong decision. If its claims hold, companies could shift routing, scoring, fraud checks and verification to a faster, cheaper model while keeping frontier LLMs for complex reasoning. That could lower AI costs while increasing the number of decisions businesses can automate.

KodeKloud gives you 1,280+ hands-on labs where you provision Kubernetes clusters, write Terraform configs, build CI/CD pipelines, configure Linux systems, containerize apps with Docker, automate with Ansible, and manage Git workflows. 

78+ playgrounds let you experiment freely in sandbox AWS environments, Kubernetes clusters, and CI/CD systems without risk. 

190+ courses across DevOps, Cloud, and AI pair theory with hands-on labs at every step. 

KodeKloud Engineer and 100 Day Challenges provide real-world job scenarios with automated grading that confirms your solutions work.

Stuck? The 55,000+ member Discord community connects you with peers and instructors ready to help. 

Every lab runs in a live environment. You deploy, you troubleshoot, you learn. No videos without context. No simulations. The kind of practice that actually builds confidence because you've done real work, not watched someone else do it.

Start Learning For Free

Evolving AI: Google's upgraded Gemini voice models can now also think through hard problems while they talk.

Google says Extended Thinking ranks first on the Artificial Analysis Speech to Speech Quality Index with 82.6 and leads the τ-Voice task test. Google AI Pro and Ultra subscribers get it in Docs while every Google AI subscriber gets it in Gmail and Keep. Developers can already build with both models through the Gemini API and inside Google AI Studio. All audio the two models create carries an invisible SynthID watermark that marks it as AI-made.

Gemini Live first showed up in August 2024, and back then it only spoke English to Android users and now 97 languages within same window/chat. It can also now talk and work on hard issues, closer to how people operate. As it becomes more natural and adaptation increases, with training and refinement in each new generation. This will be adopted into more mainstream customer facing business roles, such as support. Which eventually cuts down the mundane roles with faster rerouting to solution, while individuals orchestrating swarm of bots as specialists. Soon it will get hard to tell whether the voice on the other end of a call belongs to a real person or to a machine.

Evolving AI: Mark Zuckerberg says AI labs should set their own pace on safety rather than pause progress together.

Mark Zuckerberg is pushing back on calls for an industry-wide AI slowdown, saying each lab should move at the pace needed to train safely. He points to Meta delaying Muse for months to strengthen safety and security, using independent evaluators, and committing most compute to serving people rather than recursive self-improvement. His bet is that trust and alignment will become competitive advantages.

Zuckerberg’s position puts market incentives and company-level responsibility at the center of AI safety. The open question is whether those incentives are enough when labs also face pressure to ship faster and win users. For businesses adopting agents, the practical test will be about safety promises but more about independent evaluations, failure rates, auditability, and whether systems reliably follow user intent.

QUICK HITS

🔀 Google quietly opened Claude Opus 5 to all its engineers via Antigravity, reversing a rule that forced staff onto Gemini.

🔌 Intel spinout Cornelis raised $205M and partnered with Qualcomm on an open networking architecture aimed at Nvidia's NVLink and InfiniBand lock-in.

🖨️ A Russian attacker used AI agents to breach 395 organizations through PaperCut flaws, hitting one US high school from entry to domain admin in seven minutes.

🇨🇳 Shanghai AI Lab released Atria Dawn Preview, a 744B open-weight agentic model under MIT license, with no blog post or pricing..

 *partner link