The Week AI Went Crazy…and Then Crazier

Last week in AI was absolute chaos, in the best way possible.

A flagship model launch kicked things off, alongside two open-weight releases, a new world simulator, and an image-and-video generator with an NSFW ‘Spicy’ mode. A music model was cleared for commercial use, a one-million-token API dropped, and Tesla hinted at a step forward in autonomy while quietly sunsetting a supercomputer. 

OpenAI’s GPT-5 sat at the eye of the storm, but there was plenty happening elsewhere. Anthropic pushed coding depth with Opus 4.1, Google DeepMind advanced simulation towards AGI with Genie 3 and xAI’s Grok Imagine lit up creator circles. Alibaba introduced Qwen-Image for text-faithful image generation and rolled out its own one-million-token Flash APIs. Moreover, ElevenLabs Music arrived with licensing partners in place.

Below is the week, stitched together without loose ends.

GPT-5: The Model That Thinks… But Not Quite the God-Mode We Hoped For

OpenAI rolled out GPT-5 last week, a layered, agent-like system that knows when to “think” harder, routes tasks between fast and deep reasoning modes, and cuts factual errors by up to 45% compared with GPT-4o. Benchmarks? Off the charts in some places. Use cases? From one-prompt full-stack apps to multilingual health queries. 

And yet, the launch hasn’t landed like a moonshot.

On the ARC-AGI benchmark, Grok-4 still sits miles ahead. Reddit’s top comment summed it up as “a cost-savings play, not a frontier leap”. Gary Marcus said it was “not a giant leap”. Some developers complain it feels colder than GPT-4.5, and several users demanded the old model back, compelling Sam Altman to respond. In his weekend update, he admitted OpenAI “underestimated how much some of the things that people liked in GPT-4o matter to them”, promised warmer responses, more customisation—from emoji lovers to cold logic fans—and hinted at severe capacity crunches next week.

So yes, GPT-5 is smarter, steadier and more accessible than ever. But for many, it’s merely an update, not the God-mode model they were waiting for. Whether you love or loathe GPT-5, the below meme nails the sentiment perfectly. 

Three days before GPT-5 dropped, OpenAI pulled a move it hadn’t made since 2019—flinging open the vault on its own models. Out came gpt-oss-120b and gpt-oss-20b, released under Apache-2.0 and tuned for agentic workflows, tool use, web search, code execution and even adjustable reasoning efforts.

The 120b is a mixture-of-experts beast: 117B parameters total, approximately 5.1B active parameters per token across 128 experts. The 20B? Around 3.6B over 32 experts. Notably, both bring 128k context, grouped multi-query attention, RoPE, and a fresh o200k_harmony tokeniser to the table.

Launch partners covered almost every corner of the ecosystem: Azure, AWS (Bedrock/SageMaker), Google Vertex, Databricks, Snowflake, vLLM, Ollama, LM Studio, Cloudflare, Together, Fireworks, Baseten, Vercel, Hugging Face, along with Windows AI Foundry and Qualcomm for on-device runs. Cerebras boasted roughly 3,000 tokens per second on its wafer-scale engine (WSE), while Qualcomm pitched on-device chain-of-thought as the next big privacy or latency unlock.

The reaction split right sharply. Reid Hoffman called it proof that US labs can still match China’s open-source surge—think DeepSeek V3, Kimi K2, Qwen3 and GLM-4.5. Yet, early testers weren’t pulling punches, describing gpt-oss-120b as “overfit to reasoning benchmarks” and, against top Qwen releases, “a hallucination machine”.

Meanwhile, Alibaba’s Qwen-Image went straight for one of diffusion’s longest-running headaches: getting text right. The new 20B dense model nails multi-line layouts, paragraph-level semantics, precise edits, style transfers, object tweaks and background swaps, without breaking visual realism.

On the code and context front, Alibaba dropped Qwen Flash APIs for Qwen3-Coder and Qwen-3-2507, each packing a jaw-dropping one-million-token context window. That’s not just “long context”, that’s a context shift for entire workflows: code review, legal doc parsing, RAG pipelines. Even Qwen-Plus-Latest now plays in the same sandbox.

Tesla confirmed it’s killing Dojo, shifting training to external partners, even as it builds Cortex in Austin—a cluster powered by more than 100k NVIDIA H100/H200 GPUs—and inked a $16.5 billion deal with Samsung for the manufacturing of its AI6/Hardware 6 chips through 2033. Musk teased a 10x-parameter FSD network with better video compression and hinted at a late-August roll-out if tests hold—still supervised, still not autonomy, but consistent with the trend, bigger end-to-end vision models, more data and faster training. Then There Was…

Then came OpenAI’s livestream chart fiasco. Bars that didn’t match numbers turned into a meme—a solid reminder that trust is part of the product itself. “A mega chart screwup from us earlier,” Altman said, while the corrected figures landed in the blog. Meanwhile, Gary Marcus had a field day. Polymarket cooled. And yet, usage spiked, the free tier got reasoning, and the Reddit AMA, warts and all, answered what the keynote didn’t.

Then came the cleanest flex of the week. Responding to Elon Musk’s post stating “OpenAI is going to eat Microsoft alive”, Satya Nadella said on X, “People have been trying for 50 years, and that’s the fun of it! Each day you learn something new and innovate, partner, and compete. Excited for Grok 4 on Azure and looking forward to Grok 5!”

Translation: Microsoft is the arms dealer in this war. If OpenAI wins, Azure wins. If xAI wins, Azure still wins.

Siri is slated to get a significant upgrade: Next spring, Apple is set to roll out App Intents—enabling complex, cross-app voice commands, like editing a photo and sending it to a contact in one go—marking its boldest move yet to close the gap with AI-powered assistants like Google’s Gemini. Don’t miss: “It’s our second-largest market after the US, and may well become our largest…We’re working with local partners to make AI more affordable,” Sam Altman said regarding India. Meanwhile, a little Perplexity × Zerodha spark lit up on X: after a public nudge to bring Indian market data into Comet, Perplexity CEO Aravind Srinivas asked Nikhil Kamath, “Should we?” “Absolutely, setting up a call for Monday,” Kamath replied. If this lands, the first great AI-native finance browser may speak fluent NSE/BSE by default.

If you found this newsletter insightful, share it with a friend, a colleague, or that one person who still thinks coming around weeks like this in AI is no big deal.

If you haven’t subscribed to AIM Tv yet, now’s the time. We break down the world of AI in real time, crisp, clear and always ahead of the curve. While you are at it, here’s a quick look at some of the top stories of the week. 

• If you think PostgreSQL can handle the next wave of agentic AI, Oracle disagrees. Its new Globally Distributed Exadata Database on Exascale Infrastructure promises zero-to-hyperscale scaling, active-active replication in under three seconds and built-in vector search, positioning itself as a feature-rich superset for AI, OLTP and analytics in one unified system. Read more here. 

• When it comes to AI-driven layoffs, Microsoft calls it a transformation, TCS calls it “skill mismatch”. However, the truth is the same: AI is rewriting the rules of Indian IT’s headcount model, and denial may be costing the industry more than the cuts themselves. Read the full story here. 

• In Indian IT, AI may be the new growth story—but the biggest windfall isn’t landing in developer paychecks. From HCLTech’s ₹154 crore CEO package to mid-tier bosses pulling in over ₹100 crore, executive salaries are soaring even as employee hikes stagnate or shrink. The gap between the boardroom and the cubicle has never looked wider. Full story here. 

Now, let’s explore some exciting collaborations and exclusive insights from the AIM ecosystem, brought to you with a unique twist outside our standard editorial content. 

• Hostinger Horizons just turned vibe coding into full-blown business-building, now letting anyone launch a complete e-commerce store in minutes, no code required. From product listings to payments, shipping and discounts, the AI handles it all while you stay in creative mode. And yes, AIM readers get an extra 10% off with code ‘AIMHOSTINGER’.

• Recently concluded in Bengaluru, ABBYY’s AI Pulse Hackathon showcased how combining ABBYY Vantage’s document intelligence with agentic AI stacks like Gemini, ChatGPT, and LangChain is pushing automation from a back-office function to a front-line business advantage. Read more here. 

From Pilots to Platforms: Why GBS Needs a Rethink

Most GBS and GCC teams are still stuck in siloed AI experiments, lacking the unified platforms needed to scale. A new EdgeVerve–SSON report explains why moving beyond fragmented tools to an integrated, agentic AI approach is key to driving enterprise-wide transformation. Read the full report here.

[Webinar Alert] Agentic AI, Data Management & GCCs

AIM Research is hosting the ‘Agentic AI, Data Management & GCCs’ webinar on August 6 (India Edition) and August 7 (US Edition), exclusively for CXOs, industry leaders and AI and ML practitioners. Register now.

[Must Watch] AIM tech journalist Sanjana Gupta takes you inside India’s most powerful quantum computer at QpiAI’s Bengaluru lab, a rare, behind-the-scenes tour with founder and CEO Nagendra Nagaraja, exploring how India is building its quantum future.

Until next time,

Amit Raja Naik