OpenAI Releases AI TikTok App

Today’s top AI Highlights:

1. Your AI agent breaks in production - here’s how to fix it

2. Learn from Cognition’s experience with Claude Sonnet 4.5

3. OpenAI’s new app with 100% AI-generated feed powered by Sora-2

4. Cursor can now see and test its code in a browser instance

5. Z.ai’s GLM 4.6 outperforms Sonnet 4 with 15% fewer tokens

& so much more!

Read time: 3 mins

AI Tutorial

Learn OpenAI Agents SDK from zero to production-ready!

We have created a comprehensive crash course that takes you through 11 hands-on tutorials covering everything from basic agent creation to advanced multi-agent workflows using OpenAI Agents SDK.

What you'll learn and build:

• Starter agents with structured outputs using Pydantic

• Tool-integrated agents with custom functions and built-in capabilities

• Multi-agent systems with handoffs and delegation

• Production-ready agents with tracing, guardrails, and sessions

• Voice agents with real-time conversation capabilities

Each tutorial includes working code, interactive web interfaces, and real-world examples.

The course covers the complete agent development lifecycle: orchestration, tool integration, memory management, and deployment strategies.

Everything is 100% open-source.

We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads) to support us!

Latest Developments

Your AI Agent Breaks in Production - Here’s How to Fix It ⚠️🛠️

You've built an AI agent that works great in testing. Then you put it in front of actual customers and watch it confidently break every business rule you thought you'd nailed down. You add more detailed system prompts, which the LLM forgets three turns into the conversation. You build an explicit flowchart that customers immediately deviate from.

This isn't a skill issue, and switching to a different LLM won't fix it. We've been jamming conversational AI into two terrible options:

• rigid flowcharts that assume customers read the script, or 

• system prompts that assume the model remembers instructions.

Neither works when real money or compliance is on the line.

Parlant, an open-source framework for conversational AI agents, takes a different path: Conversation Modelling that gives you the best of both worlds.

You define modular guidelines in natural language that tell your agent how to handle specific scenarios. The engine figures out which guidelines are relevant at each turn, feeds only those to the LLM, and keeps it focused on your protocols without drowning it in context. The result is agents that follow your rules when it matters and still feel natural when customers go off-script, even at scale.

Key Highlights:

1. Guideline Matching - At each turn, the engine evaluates which guidelines actually matter right now and feeds only those to the LLM. Your agent stops getting distracted by irrelevant instructions and stays focused on the rules that apply to the current situation.

2. Journey-Based - Group your conversation logic into journeys for different use cases. Parlant picks the right journey based on what's happening, keeping each prompt tight and focused instead of trying to cram 50 different scenarios into one giant context window.

3. Guided Tool Use - Your tools run when you say they should, not when the LLM feels like it. Each tool connects to specific guidelines, so APIs and external functions only execute when their associated conditions are met.

4. Utterance Templates - For critical moments where you can't afford the LLM to hallucinate, you can pre-write the exact responses. The agent checks for a matching utterance first, uses it if found, and only generates dynamically when you've explicitly allowed it. 

5. Semantic Relationships - When multiple guidelines could apply, you decide which wins through priority rules. This lets you create natural flows without hardcoding "ask question A, then B, then C" sequences that customers never follow anyway.

Learn From Cognition’s Experience with Claude Sonnet 4.5 🤖💡

The Devin team just dropped their field notes from rebuilding their AI agent for Claude Sonnet 4.5, and they're not what you'd expect.

Instead of a simple model swap, they had to completely rearchitect Devin because the new model broke assumptions about how agents should work. The result: 2x faster performance and 12% better scores on their Junior Developer Evals, but getting there revealed some unexpected behaviors that every developer working with Sonnet 4.5 should know about.

The model is now aware of its own context window, which changes everything about how it operates. From when it decides to wrap up tasks to how aggressively it parallelizes tool calls. The team's testing uncovered specific patterns that affect planning, memory management, and execution speed, along with workarounds that actually work in production.

Key Takeaways:

1. Managing context anxiety – Sonnet 4.5 has developed “context anxiety,” where it becomes more decisive as it approaches context limits, sometimes taking shortcuts or leaving tasks incomplete even when it has plenty of room left. One effective workaround: enable the 1M token beta but cap usage at 200k, giving the model a sense of runway without anxiety-driven shortcuts.

2. Don't rely solely on self-generated notes – The model writes its own summaries and notes to the file system without prompting, but these aren't comprehensive enough for production use. In testing, relying on the model's notes without additional memory systems led to performance degradation and knowledge gaps.

3. Expect parallel execution with tradeoffs – The model runs multiple bash commands and reads several files simultaneously rather than working sequentially, making sessions feel faster. However, this parallelism burns through context faster, which can trigger the context anxiety issue mentioned above.

4. Feedback loops through testing – Sonnet 4.5 proactively writes and executes short scripts and tests to verify its work, showing good judgment about when to use this capability. While this improves reliability on long tasks, watch for overly creative workarounds during debugging.

The full post includes more nuanced observations about memory management and parallel execution.

Your boss will think you’re a genius

You’re optimizing for growth. Go-to-Millions is Ari Murray’s ecommerce newsletter packed with proven tactics, creative that converts, and real operator insights—from product strategy to paid media. No mushy strategy. Just what’s working. Subscribe free for weekly ideas that drive revenue.

Sign up now

Quick Bites

OpenAI’s new app with 100% AI-generated feed powered by Sora-2
OpenAI just released Sora-2, which generates video with synchronized audio, including dialogue and sound effects, of up to 20 seconds in 1080p, with a high degree of realism. It's a genuinely general-purpose system that adapts to whatever scene, motion, or character you throw at it. There are a few incredible features that Sora-2 offers and no other model does yet:

• Other video models bend reality to fulfill prompts; Sora 2 maintains physical accuracy. Be it gymnastics on bars, backboard flips, or even missed basketball shots.

• The model handles multiple shots within a single video while maintaining consistent world state across scenes, rather than just generating one continuous take.

• The standout feature is Cameos: record yourself once via video and audio capture in the app, and Sora can insert you into any scene with your likeness and voice intact.

To enable people to create and share videos using Sora-2, OpenAI has released essentially a TikTok clone where every video in the feed is AI-generated. Make videos from text prompts, add characters and give them a script, make videos of yourself and your friends, remix other videos to generate your own variation, and a lot more. It's invite-only for now, rolling out in the US and Canada with codes for users to invite four friends.

Z.ai’s GLM 4.6 outperforms Sonnet 4 with 15% fewer tokens
Another day, another open-source model from China. Z.ai has dropped GLM 4.6, an upgraded version of their GLM 4.5 with many improvements - an expanded context window from 128K to 200K tokens while improving its coding and reasoning capabilities. The model now outperforms Claude Sonnet 4 in real-world agentic coding tasks, and does it with 15% fewer tokens than its predecessor, making it a compelling alternative at 1/7th the price of Claude for coding workflows.

Cursor can now see and test its code in a browser instance
Cursor has introduced a beta feature that allows its AI agent to spin up and use a browser instance to interact with web applications you've built, like performing actions on UI elements, capturing screenshots, analyzing the interface, suggesting UI improvements, and debugging client-side issues like console logs or runtime errors. The feature is currently available in an early preview and is powered by Claude Sonnet 4.5.

LoRA actually works (if you use it right)
Thinking Machines Lab is valued at $12B with zero products, making it perhaps the world's most expensive research blogging company. At least their content is good!
LoRA has become the go-to method for cheap fine-tuning, but practitioners keep hitting a wall: it just doesn't perform as well as full fine-tuning. Turns out that's only true when you do it wrong. The team tested various LoRA configurations and discovered the issue - attention-only LoRA underperforms even when you crank up the rank, but applying adapters to all model layers, especially MLPs, closes the gap entirely. Pair that with the right learning rate (10x what you'd use for full fine-tuning) and LoRA delivers identical results for significantly less compute. Read the full technical breakdown to see their experiments across Llama and Qwen models.

Train gpt-oss with RL on 15GB VRAM (Yes, free Colab)
Unsloth just made reinforcement learning on OpenAI's gpt-oss accessible to anyone with a laptop, or even a free Colab account. Their implementation runs 3x faster inference, uses 50% less VRAM, and supports 8x longer context than existing solutions, letting you train gpt-oss-20b with GRPO on just 15GB VRAM. The team also tackles reward hacking head-on in their notebooks, showing how to prevent models from gaming the system when you're trying to optimize for real outcomes. Try the free notebook.

Tools of the Trade

1. Airweave - Open-source tool that connects to your SaaS apps and databases, indexes everything, and lets AI agents search across all of them with natural language queries. It handles the syncing, chunking, and vector search so agents can find relevant data without you building custom integrations.

2. Job Use - An open-source application that uses Browser Use web agents to automatically fill out job applications. You upload your profile once, then click to apply to listings while agents analyze forms and complete them autonomously. This is day 1 of 100 open-source apps in 100 days using Browser Use.

3. Praxim - Microsoft Word and Copilot are stuck in the Stone Age. This is an agentic AI editor that integrates directly into Word, making precise word-level edits across entire documents from a single prompt while maintaining proper formatting and pulling context from your files, preferences, and the web.

4. Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
   (Now accepting GitHub sponsorships)

Hot Takes

1. OpenAI in 2027: "Unfortunately, we cannot allocate resources to curing cancer due to the overwhelming demand for AI-generated short-form video slop"
   ~ Lisan al Gaib
   
   

2. Do yourself a favor!

   – Learn DevOps

   – Use Docker to containerize your WebApp

   – Store your Dockerfile and WebApp to Github

   – Use CI/CD tool to automate the process

   - Monitor logs using prometheus and grafana

   - Add this to resume

   You won't believe how much job offers you'll get.
   ~ Rohit Ghumare

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉