OpenAI Pivots to ‘Modular GPT‐5’, Stargate Project Stalls

Good morning. It’s Wednesday, July twenty third.

On today in tech history: In 1972Terry Winograd released SHRDLU, an AI that would understand and manipulate a virtual block world via typed language. A landmark in symbolic AI, it combined a semantic parser, planner, and reasoning system to take care of a dynamic model of its environment and reason about spatial relationships, thereby bridging language, vision, and planning.

OpenAI pivots to modular GPT‑5 as Stargate project stalls
DeepMind’s Gemini clinches AI’s first math gold as controversy clouds celebrations
xAI’s Grok 4 triples iOS revenue; next-gen Colossus 2 supercluster nears launch
5 Latest AI Tools
Latest AI Research Papers

You read. We listen. Tell us what you’re thinking that by replying to this email.

Your agent failed. Here’s why.

Debugging agents often means reading through traces and attempting to piece together what went improper. It’s slow, manual, and infrequently leaves teams unsure of the particular issues.

Atla helps by robotically detecting where and why failures occur, summarizing traces so that you don’t need to dig through every step, and showing error patterns across runs. The platform also provides actionable suggestions for improvement, and allows teams to check prompt changes and compare results side-by-side.

Atla is built upon proprietary research on evaluating agents, and is designed to exchange vibe checks with grounded evaluation.

The platform is open for early access–try it at no cost and improve your agents.

^{Thanks for supporting our sponsors!}

Today’s trending AI news stories

OpenAI pivots to modular GPT‑5 as Stargate project stalls

OpenAI is reportedly near launching a dynamic “router” for ChatGPT that may robotically select one of the best AI model for every user prompt. Leaked posts from OpenAI insiders hint the router won’t kill manual control but will quietly switch to stronger models for complex or critical queries, like a real-time triage between reasoning, creative, and gear‑using subsystems. Rumors suggest this is an element of a move toward GPT‑5 as a modular, task-aware system somewhat than a single monolithic model.

Open is ready to pay Oracle $30B/12 months for an enormous 4.5GW Stargate data center. Image: Sam Altman on 𝕏

On the infrastructure side, SoftBank and OpenAI’s ambitious $500 billion Stargate project to construct massive AI data centers has stalled just six months after launch. After disagreements over land use and strategy, plans have scaled back to a single Ohio facility by 12 months’s end. SoftBank raised $3 billion, but OpenAI has pivoted, securing a $30 billion cope with Oracle to construct a Texas data center housing 400,000 Nvidia GB200 chips and partnering individually with CoreWeave.

ChatGPT is now fielding over 2.5 billion prompts day by day, adding up to just about a trillion a 12 months. Of those, about 330 million requests come from U.S. users alone, OpenAI confirmed to The Verge. The firm’s first economic report, led by Chief Economist Ronnie Chatterji, shows ChatGPT driving sharp productivity gains. Despite this explosive growth, JPMorgan cautions OpenAI’s moat is “increasingly fragile,” citing rising litigation, price battle, and dependence on consumer subscriptions. Read more

DeepMind’s Gemini clinches AI’s first math gold as controversy clouds celebrations

Google DeepMind’s Gemini has set a brand new milestone in AI reasoning, earning a gold medal–level rating on the 2025 International Mathematical Olympiad by solving five of six complex problems using only natural language.

This historic feat makes Gemini the primary AI to be officially graded at gold level, outperforming last 12 months’s silver-winning AlphaGeometry. Its success stems from a “parallel pondering” architecture that explores multiple solution paths concurrently, combined with advanced reinforcement learning and curated mathematical proof data.

A sophisticated version of Gemini with Deep Think has officially achieved gold medal-level performance on the International Mathematical Olympiad. 🥇

It solved 5️⃣ out of 6️⃣ exceptionally difficult problems, involving algebra, combinatorics, geometry and number theory. Here’s how 🧵

— Google DeepMind (@GoogleDeepMind)
4:32 PM • Jul 21, 2025

Btw as an aside, we didn’t announce on Friday because we respected the IMO Board’s original request that each one AI labs share their results only after the official results had been verified by independent experts & the scholars had rightly received the acclamation they deserved

– Demis Hassabus (@demishabisabis)
4:47 PM • Jul 21, 2025

Alongside this, Google released Gemini 2.5 Flash-Lite, its fastest, most cost-efficient model yet, designed for scaled production use with a big context window and native reasoning tools. Early adopters have reported significant latency reductions and energy savings.

Complementing this, Gemini 2.5 now supports conversational image segmentation: users can discover objects or abstract concepts like “damage” or “clutter” using natural language, even reading on-screen text. Results return as JSON with pixel masks and coordinates, all accessible through the Gemini API and Google AI Studio.

These wins also reignite an old but practical query: is intelligence grounded in rigid symbols, or do symbols simply bend to thought? By reaching gold‑level IMO scores through pure natural language reasoning, i.e., no formal solvers, DeepMind and OpenAI’s models challenge the concept true understanding have to be symbolic.

DeepMind’s Andrew Lampinen frames symbols as scaffolding for intuition, echoing Wittgenstein: meaning emerges through use, not structure alone. Hybrid neuro‑symbolic systems like AlphaProof once led this field, but these results show deep learning alone can now rival human mathematical reasoning. Read more.

xAI’s Grok 4 triples iOS revenue; next-gen Colossus 2 supercluster nears launch

Elon Musk’s xAI is scaling at an aggressive pace on each infrastructure and product fronts. The corporate is finalizing Colossus 2, an enormous latest supercomputing cluster slated to deploy over 550,000 NVIDIA GB200 and GB300 GPUs for AI training inside weeks. This follows Colossus 1, which already runs 230,000 GPUs, including 30,000 GB200s, to coach the Grok family of models, while inference workloads remain on external cloud providers. NVIDIA CEO Jensen Huang has described xAI’s build-out speed as unmatched within the industry.

xAI saw an enormous revenue spike after launching its latest model, Grok 4. Released on July 9, Grok 4 pushed day by day iOS revenue from $99,000 to $419,000 by July 11, a jump of 325%, in keeping with Appfigures. Downloads also surged 279% to just about 197,000. The momentum continued for days before dipping barely. Per week later, xAI added raunchy AI companions for “Super Grok” subscribers at $30/month. While these drove downloads up 40% to 171,000, revenue only rose by 9%, showing less impact than the model launch itself.

Despite the steep price, well above many competitors, demand was strong enough to briefly lift Grok to No. 3 overall within the U.S. App Store. Read more.