TLDR AI 2025-07-22

The future of work will not be on laptops and phones. Immersed is an AI + AR/VR company shows a glimpse a into the future: a virtual workplace that's become the #1 productivity app on Meta's Quest Store.

1.4M professionals, including Fortune 500 teams, now use Immersed's Spatial Computing technology as their primary work environment. Meta, Google, Microsoft, and Qualcomm are already on board as strategic partners or backers.

Now, they're ready to ship their own headsets to 75K+ people on the waitlist, and $71M in projected first year sales — and are raising an equity crowdfunding round to fuel growth.

Invest by July 30th for an 8% discount

Alibaba's Qwen team has released an updated Qwen3-235B-A22B, a non-reasoning model, with major improvements.

Six months after the White House announcement, the venture has completed zero data center deals as SoftBank and OpenAI clash over building sites tied to SB Energy, a SoftBank-backed developer. While the joint venture targets a single small Ohio facility by year-end, OpenAI has independently secured a $30 billion annual Oracle deal for 4.5 gigawatts—nearly matching Stargate's entire first-year capacity goals.

Hot on the heels of OpenAI's gold medal on Saturday, Google announced its model solved five of six IMO problems, earning official gold-medal certification from tournament coordinators. The breakthrough marks a shift from Google's previous approach—last year's AlphaProof required expert translation into formal mathematical languages and days of computation. Deep Think operates entirely in natural language, exploring multiple solution paths in parallel.

Extracting information from a complex PDF can be a pain. The industry has spent enormous amounts of effort and money on OCR, layout detection, and parsing pipelines, just to still lose the information that matters most. Vision Language Models have gotten good enough to understand documents directly without any parsing, OCR, or reconstruction. Instead of several fragile steps, you can now have one robust operation that preserves every chart, table relationship, and visual cue.

The first nation to achieve AGI will gain a decisive "fourth offset"—a military advantage comparable to nuclear weapons or precision-guided munitions. The analysis estimates that a foundation model trained with 2e29 FLOPs would automate AI research itself, producing output equivalent to millions of scientists. With the US projected to reach this threshold by 2030 while China faces chip constraints, the author advocates aggressive investments in power, security, and manufacturing to secure a first-mover advantage.

Training trillion-parameter models efficiently requires optimizers that learn more from each token, but these advanced methods often cause training crashes at scale. Kimi K2 solves this with MuonClip, combining the token-efficient Muon optimizer with a novel QK-Clip technique that prevents attention weights from breaking during training.

Prima Mente's Pleiades is a family of epigenetic foundation models (90M–7B parameters) trained on 1.9T tokens of human methylation and genomic data. The models outperform DNA-only baselines in cfDNA generation, neurodegenerative disease detection, and epigenomic prediction by integrating methylation context with genomic sequences via hierarchical attention and coordinate-aware embeddings. Early results show promising diagnostic accuracy for Alzheimer's and Parkinson's, establishing a foundation for multimodal brain models and biomarker discovery.

Apple has published a tech report that details how its models were trained, optimized, and evaluated. It walks through multiple aspects of the company's new models, including their architecture, data sources, pre-training, post-training, tool use development, optimizations, and benchmarks. The paper is very technical - this post presents a few particularly interesting highlights.

OpenAI claims that ChatGPT receives 2.5 billion prompts from all over the world every day. Google receives 5 trillion queries per year, an average of just under 14 billion daily searches. ChatGPT has shown impressively rapid growth. OpenAI CEO Sam Altman said in December that users sent over 1 billion queries a day, so the company's search volume has more than doubled in around eight months.

While there was some impact from the companions launch, it was decidedly smaller than the launch of Grok 4.

Former Meta executive and soon-to-be former Instacart CEO Fidji Simo outlined an optimistic vision of AI democratizing access to knowledge, healthcare, creative tools, economic opportunities, and personalized support in a memo to OpenAI staff.

Meta and AWS have partnered to support 30 U.S.-based startups through a six-month program that will offer compute credits, engineering mentorship, and technical support for startups building generative AI applications with Llama models.