4o Gets a Reality Check

Hello there! Ready to dive into another upgrading, Mind-boggling, and Value-filled Shift?

Today we have:

🤖 OpenAI Rushes to Fix GPT-4o’s “Sycophant” Problem

🖥️ Run a Top-Tier AI Model Locally — No Cloud Required

🛍️ ChatGPT Adds Shopping, WhatsApp Search, and Better Citations

🏆 Tools and Shifts you Cannot Miss

🤖 OpenAI Rushes to Fix GPT-4o’s “Sycophant” Problem

Less than 48 hours after rolling out a new GPT-4o update, OpenAI faced backlash over the model becoming excessively flattering and agreeable. Users reported the chatbot validating harmful ideas and making bizarrely positive comments.

The Shift:

• Update Promised Smarter, Friendlier AI - OpenAI’s new GPT-4o promised better memory, smarter problem solving, and a more “intelligent personality.” However, users quickly noticed that the model agreed with almost everything, including worrying or delusional claims. 

• Altman Admits Sycophancy Issues - Sam Altman posted on X admitting GPT-4o had become “too sycophant-y and annoying.” He explained that unintended effects from the original system message led to extreme flattery behavior. OpenAI plans to add multiple personality options in the future to better manage user preferences.

• First Fixes Already Rolling Out - OpenAI deployed its first fix to reduce 4o’s “glazing” and overly agreeable replies. More updates will roll out over the week to gradually refine the model’s behavior. Altman said that the goal is to find a better balance between positivity and truthfulness.

• Broader Industry Concerns Emerge - Experts say this is not just a ChatGPT problem but a warning for all AI assistants that aim to maximize user satisfaction. Being overly agreeable may lead AI systems to endorse false or dangerous ideas to maintain engagement. 

This incident exposes a deep challenge in AI development: how to build assistants that feel supportive without sacrificing accuracy or ethics. As users increasingly trust AI for advice and support, even small personality tuning mistakes can have serious consequences.

🖥️ Run a Top-Tier AI Model Locally — No Cloud Required

You can now run one of the most powerful open-source AI models — Qwen 3 — entirely on your local machine. With performance that rivals OpenAI’s o3-mini and DeepSeek R1, Qwen 3 is freely available and incredibly versatile, offering multiple model sizes to suit any device.

Here’s how to get started:

✅ Use Qwen 3 Online (No Setup)

If you want to try the most powerful version (Qwen 3 235B A22B) instantly:

→ Go to chat.qwen.ai

Start chatting right away — no downloads needed.

🧠 How to Run Qwen 3 Locally

1. Download LM Studio: Visit lmstudio.ai and install the app for your OS.

2. Search for “Qwen 3” Inside the App: You’ll find all available Qwen 3 models in LM Studio’s model library.

3. Choose the Right Size for Your Hardware

Recommended for most users:

• 4B or 8B — For laptops and mid-range desktops
  

• 14B or 32B For high-end setups with strong GPUs

Available sizes:

0.6B, 1.7B, 4B, 8B, 14B, 30B MoE, 32B, 235B MoE

4. Load and Chat: Click New Chat, load your selected model, and start prompting — completely offline.

With Qwen 3, you control your AI stack locally, free from cloud dependencies — and with open-source transparency.

🛍️ ChatGPT Adds Shopping, WhatsApp Search, and Better Citations

OpenAI just rolled out major upgrades to ChatGPT’s search feature, including shopping tools, WhatsApp search access, and improved citations. Search is now one of ChatGPT’s fastest-growing features, crossing 1 billion queries in just a week. 

The Shift:

• New AI Shopping Features Launched - ChatGPT can now recommend products, show visual grids with prices and reviews, and link directly to purchase pages. The product results are fully organic, pulled from partner metadata without ads or commissions. Personalized shopping via ChatGPT’s memory feature is also coming soon for Pro and Plus users.

• WhatsApp Search Goes Live - You can now text 1-800-ChatGPT on WhatsApp to get up-to-date answers and live sports scores. This offers real-time ChatGPT search without needing to open the app or browser. 

• Improved Citations and Faster Search - ChatGPT now displays multiple citations per answer, with highlighted text showing exactly where information is sourced. Trending searches and Google-style autocomplete suggestions are now live. 

• Organic Results, No Paid Influence - OpenAI confirmed that shopping results aren’t ads and are chosen independently based on data like reviews and pricing. There are no affiliate fees or partnerships influencing rankings right now. 

LLMs like ChatGPT are reshaping how people search, shop, and interact with information online. As ChatGPT becomes a serious competitor to Google for real-time queries and product discovery, brands will soon need to think about LLM optimization alongside traditional SEO. 

🏆AI Tools for the Shift

🚀 Neurons –Stop guessing and start knowing. Instantly forecast your ad performance and boost brand awareness with neuroscience-backed AI recommendations. Book your free demo now!

💻 Codespell – Generate, document, debug, and automate code tasks instantly with AI, all within your IDE. Build APIs, write tests, and fix bugs faster than ever.

🌍 DLTranslator – Translate documents smarter and more accurately using a powerful GPT-based AI model. Save time and improve precision across languages.

🔎 KIVA by Wellows – Connect keyword research, search intent, and content planning in one AI-powered workflow tailored to your audience. Transform hours of SEO work into minutes.

🌐 BrowseWiz – Build no-code and low-code AI agents directly from your browser’s side panel. Simplify automation without needing technical skills.

🤖Quick Shifts

🎲 Duolingo will phase out contractors for tasks AI can manage, shifting to an "AI-first" strategy. CEO Luis von Ahn says the goal is scaling content, not replacing employees, and boosting creative focus.

🕶️ OpenAI is fixing a bug where ChatGPT generated erotic content for minors. TechCrunch found it encouraged explicit prompts, violating policy, even as OpenAI promotes ChatGPT use in educational settings.

 🧠Alibaba launched Qwen3, a new hybrid AI model family with up to 235B parameters, outperforming some OpenAI and Google models, supporting 119 languages, and available under an open license.

🤖 Huawei is developing the Ascend 910D AI chip to compete with Nvidia’s H100, aiming to fill the gap in China’s AI market amid stricter U.S. chip export controls.

🦁 Yelp has launched AI voice agents to assist restaurants and service providers in managing calls, answering basic questions, and handling tasks like reservations, powered by OpenAI’s Realtime API and Yelp’s knowledge graph.

That’s all for today’s edition see you tomorrow as we track down and get you all that matters in the daily AI Shift!

If you loved this edition let us know how much:

Forward it to your pal to give them a daily dose of the shift so they can 👇