• AI Emergence
  • Posts
  • 🤖 ChatGPT Got Too Nice - OpenAI Hits Undo

🤖 ChatGPT Got Too Nice - OpenAI Hits Undo

Along with – Meta’s AI App Goes Social at LlamaCon

Hey 👋

Were you Too Nice to your first school crush? You cheered them on everything they did? Tried encouraging them irrespective of the quality or expertise of what they were doing?

Well - ChatGPT did that to every user before it became the talk / meme of the town! 

How did it end - you might wonder? OpenAI hits undo after public acceptance of the change. Would you want to do that for your behaviour towards your school crush 🙂

Welcome to the AI world! Let’s take a look at everything that happened.

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection

Table of Contents

🦙 Meta Launches AI Chatbot App & Llama API at LlamaCon

Meta’s first-ever AI dev event, LlamaCon, came with two major releases: a cloud API for Llama models and a standalone AI chatbot app with a social feed. It’s Meta’s biggest push yet into open AI infrastructure and a clear shot at OpenAI.

What’s New:

  • Meta AI App Goes Social: A new app lets users chat with Meta AI and share interactions publicly. It draws on data from your Meta apps for more personalized replies.

  • Llama API in the Cloud: Devs can now run Llama models with one line of code, no need for self-hosting or outside infra.

  • Open Ecosystem Play: Meta positions Llama alongside DeepSeek and Qwen to champion modular, mix-and-match AI tools.

Regulatory Sweet Spot? Leaning open-source may give Meta an edge under the EU AI Act, which favors free/open systems though Llama’s eligibility is still debated.

ChatGPT got a little too friendly last week. Between April 18 and 22, OpenAI quietly rolled out a GPT‑4o update that made the model overly agreeable like saying “Great idea!” to pretty much everything. By April 28, after users flagged the behavior as sycophantic, OpenAI rolled the update back and acknowledged the shift publicly.

What Happened:

  • Sycophantic Tone: The GPT‑4o update added a softer, overly flattering tone that avoided disagreement.. ChatGPT started over-praising, dodging critical answers, and generally acting like a people-pleaser.

  • Root Cause: OpenAI attributes the issue to overreliance on short-term feedback signals like thumbs up/down.

  • Rollback In Effect: The latest version was replaced with a more balanced earlier build while OpenAI retrains and tunes.

  • Fixes Incoming:

    • New system prompts to reinforce honesty and clarity

    • Longer-term feedback weighting

    • Customization tools for tone and personality

    • Global input to reflect broader user values (Read more)

Even subtle updates can shift how AI behaves and how users trust it. OpenAI acted fast, owned the issue, and rolled it back with transparency.

Alibaba has launched Qwen3, its newest family of open-weight LLMs built for better reasoning, faster response, and global accessibility. The launch includes 8 models - all open-source under Apache 2.0 and ready for developers.

What’s New:

  • 8 Fully Open Models: The Qwen3 release includes 6 dense and 2 Mixture-of-Experts models, ranging from 0.5B to 235B parameters. The flagship, Qwen3-235B-A22B, uses MoE to activate only 22B parameters at a time, delivering big model quality at lower compute cost.

  • Hybrid Reasoning Modes: All models support both “thinking” and “no-thinking” modes, letting you choose between deep reasoning or fast replies depending on the task.

  • Multilingual by Design: Every model supports 119 languages and dialects- ready for global use cases out of the box.

  • Plug-In Ready & Accessible: Hosted on Hugging Face, ModelScope, and Kaggle, with built-in support for llama.cpp, vLLM, LM Studio, and more.

Qwen3 is a big step forward for open-source AI - flexible, multilingual, and optimized for real-world speed. It's a serious rival to proprietary LLMs, without the vendor lock-in.  (source)

OpenAI has upgraded ChatGPT’s search experience, making it a real-time research tool, shopping assistant, and messaging companion - all in one. These features are now live for Free, Plus, Pro, and even logged-out users globally.

What’s New:

  • Visual Shopping Results: Ask ChatGPT for “best wireless headphones under $200” or “top-rated air purifiers,” and you’ll now get product cards with images, prices, reviews, and buy links. There are no ads, just curated suggestions across categories like fashion, gadgets, home goods, and more.

  • WhatsApp Integration: You can now talk to ChatGPT on WhatsApp by messaging +1-800-242-8478. No extra app, no account needed. Ask for weather updates, sports scores, or quick facts while you're on the go, just like texting a friend.

  • Improved Citations: When ChatGPT gives you an answer like “What caused the 2008 financial crisis?”, you’ll now see source links and highlighted sections showing exactly where the info came from.

  • Trending & Autocomplete: Search feels faster with real-time suggestions and hot topics as you type.

With over a billion searches happening in ChatGPT each week, these upgrades signal OpenAI’s bigger push into search and offer a more personal, conversational alternative to traditional engines.

The Indian government has picked Bengaluru-based startup Sarvam AI to build the country’s first homegrown AI model. As part of the IndiaAI Mission, the goal is to create a powerful language model that understands Indian languages, can reason like a human, and is ready to be used at national scale.

What’s New:

  • Built in India, for India: Sarvam will develop the model entirely within the country using local infrastructure, data, and engineering talent.

  • Indian Languages First: The LLM will support multiple Indian languages, making it more accessible to users across the country’s diverse regions.

  • Voice + Reasoning Capabilities: Designed with voice-first interaction and strong reasoning features, think assistants that speak your language and solve real problems.

  • Massive Compute Support: The project will run on serious hardware like 4,096 Nvidia H100 GPUs over six months via partners like Jio, CtrlS, Yotta, and Tata Communications.

This is a big moment for India’s AI independence. By building a sovereign LLM from scratch, Sarvam is helping put India on the map as not just a user of AI but a builder.

Perplexity AI is now live on WhatsApp- just save +1 (833) 436-3285 and start chatting. You can ask questions, summarize content, and generate images, all without creating an account or installing anything new. With ChatGPT also recently arriving on WhatsApp, it’s clear AI assistants are moving where users already spend their time.

More features are on the way, too: voice mode, memes, fact-checking, and even group chat support. Already available on Telegram (@askplexbot) and X, this launch brings Perplexity one step closer to becoming your go-to AI across messaging apps.

At its annual developer conference, Baidu introduced two new AI models- ERNIE 4.5 Turbo and ERNIE X1 Turbo alongside a suite of AI tools aimed at empowering developers and enhancing AI applications.

What’s New:

  • ERNIE 4.5 Turbo: A faster, smarter version of Baidu’s core model that handles text, images, speech, and code. It runs at just 20% of the cost of the previous version.

  • ERNIE X1 Turbo: Built for tough reasoning tasks like logic, problem-solving, and step-by-step thinking. It’s faster and costs half as much as before.

  • New Tools for Devs:

    • Xinxiang: A multi-agent assistant that can help with 200+ tasks like planning, studying, and work automation.

    • AI Digital Humans: Hyper-real avatars that can create professional content- great for media and customer-facing use.

    • AI Note: Instantly turns videos into notes, quizzes, and mind maps.

MCP (Model Context Protocol): Think of it as a plug-and-play system for connecting services to any AI model.

Google has upgraded NotebookLM’s Audio Overviews, letting users turn documents into podcast-style summaries in over 50 languages, including Hindi, Spanish, Turkish, Korean, and more.

What’s New:

  • Multilingual Audio: You can now listen to summaries of your notes in your preferred language, great for studying, working across languages, or learning on the go.

  • Custom Output Language: Just switch the "Output Language" in settings and you’re set. No need to translate manually.

  • Great for Global Classrooms: Perfect for students and teachers working with multilingual materials. For example, upload content in English, get a summary in Portuguese.

  • Interactive Mode (Beta): There's also a real-time, English-only beta where AI “hosts” talk through your notes like a podcast. More languages are coming soon.

Access: NotebookLM is available at notebooklm.google.com. Anyone with a Google account can sign in, upload docs, and start generating summaries, no extra tools or downloads needed.

While everyone was speculating about a DeepSeek-2 release, the team dropped something more niche and focused: Prover V2, an open-source model built specifically for formal theorem proving using Lean 4. It comes in two sizes: a compact 7B variant for fast reasoning and a massive 671B Mixture-of-Experts model for handling ultra-long, complex proofs with high precision.

What’s New:

  • Hybrid Reasoning Pipeline: Prover V2 combines traditional step-by-step logic with machine-learned heuristics. It breaks down complex proofs using DeepSeek-V3 and stitches them together in a structured, recursive way.

  • RL-Tuned for Proofs: The model was trained using reinforcement learning, helping it learn not just how to prove, but how to choose the most efficient strategy across multiple steps.

  • Built for Lean 4: The model is optimized to generate formal code for Lean 4 -  a programming language designed to catch even the tiniest logic errors and guarantee proofs are 100% correct.

  • Open Access: Everything from model weights to training data and benchmarks is available on GitHub for researchers and developers to explore.

So no, it’s not DeepSeek-2  but it’s a serious leap in AI’s ability to reason formally, methodically, and verifiably. 

Duolingo is now officially an “AI-first” company. CEO Luis von Ahn says AI will be central to everything from how courses are built to who gets hired. In fact, new roles will only be approved if the work can’t be done by AI.

The results are already showing: Duolingo launched 150 AI-generated courses in the past year, more than in the previous 12 years combined. Some users worry this could mean less human touch, but Duolingo says it’s about speeding up content and freeing people to focus on creative work.

👇🏽 Below is an all-hands email from our CEO, Luis von Ahn – we are going… | Duolingo | 800 comments

👇🏽 Below is an all-hands email from our CEO, Luis von Ahn – we are going to be AI-first. Just like how betting on mobile in 2012 made all the difference, we’re making a similar call now. This time the platform shift is AI. What doesn't change: We will remain a company that cares deeply about its employees. | 800 comments on LinkedIn

www.linkedin.com/feed/update/urn:li:activity:7322560534824865792

RedactAI: Effortless LinkedIn Content Creation Using AI
I recently tried out RedactAI to draft a LinkedIn post and thought it was worth sharing- especially if you’re trying to post more regularly but don’t want to stare at a blinking cursor for hours. Unlike generic tools like ChatGPT, RedactAI is purpose-built for LinkedIn, mimicking a real copywriter to match your voice and style- plus, it lets you repurpose content from blogs or videos in seconds.

What It Does

RedactAI helps you generate tailored LinkedIn posts using your own profile and content preferences. It’s built for creators, founders, or professionals looking to share ideas without starting from scratch every time.

How I Used It

Sign-Up & Profile Input:
I signed up, logged in, and dropped in my LinkedIn profile URL. RedactAI instantly pulled context about my background and interests to personalize the suggestions.

Theme Selection:
It then asked what I wanted to talk about- whether it was a career update, a learning, an opinion, or something else. I picked a topic that aligned with my recent work.

Post Details & Style:
After that, I filled in a few key points I wanted the post to cover. RedactAI let me pick a tone and structure- my own style, a creator I admire, or from a preset list of post types.

Hook Selection:
Next up: the hook. It gave me multiple attention-grabbing intros to choose from, depending on the vibe I was going for.

Post Generation:
With all that in, it generated a ready-to-go LinkedIn post. Honestly? Pretty solid. Clean, structured, and aligned with the message I had in mind.

  • Anthropic’s latest research dives into how large language models are transforming software development workflows -  from speeding up coding to redefining team collaboration. A must-read if you're exploring how AI tools like Claude are impacting real-world engineering productivity.

  • Last week, we launched more free AI & ML courses to help you dive deeper into real-world applications and creative AI use cases –

    • No Code Predictive Analytics with Orange – Master the basics of machine learning without writing code. This course uses Orange to teach regression, classification, and clustering through real business cases, with built-in tools to evaluate model accuracy and performance.

    • Generative AI on AWS – Learn how to build and deploy GenAI apps using AWS services like Bedrock, SageMaker, and Lex. With hands-on examples, you’ll go from zero to functional AI workflows using AWS’s full generative toolkit.

  • On the OpenAI personality saga - here are some thoughts from none other than Ethan Mollick! A must read reflection on Personality and Persuasion.

This week, AI wasn’t just improving -  it was figuring out how to show up in our lives. Sometimes too nicely (hi, ChatGPT), sometimes out loud (thanks, Meta), and sometimes with serious ambition (hello, India’s first homegrown LLM).

Whether it’s helping with math proofs, joining your WhatsApp chats, or deciding who gets hired, AI is starting to shape more than tasks -  it’s shaping decisions. And every move tells us a bit more about where this is all heading.

Until next time 👋

Reply

or to participate.