Try Workbench free for 20 min/ day →

Good morning. It's Wednesday, May 6, and we're covering growing worries about AI compute shortages, a backlash over claims that chatbots may be conscious, and a breakthrough model aiming to dramatically cut the cost of scaling AI.

Know someone who'd love this? Forward it their way. First time reading? Sign up here to keep up with the future of tech.

SubQ Challenges Transformer Norms
Instead of scaling compute alongside context, Subquadratic claims its SubQ model reduces attention overhead by nearly 1,000x while extending context to 12 million tokens. The model introduces a sub-quadratic sparse-attention architecture that selectively processes relevant token relationships rather than all possible pairs, diverging from standard transformer designs. Early benchmarks cited by the company show SubQ running 52x faster than FlashAttention at 1 million tokens and operating at under 5% of the cost of Opus-class models.

OpenAI Smartphone Plans Accelerate
Rather than easing into hardware, OpenAI is reportedly targeting mass production of its first smartphone by early 2027, an unusually aggressive timeline for a new entrant. Analyst Ming-Chi Kuo says the device will compete directly with the iPhone, using a customized MediaTek Dimensity 9600 chip built on TSMC’s N2P node and featuring dual NPUs and LPDDR6 memory. Estimated shipments could reach 30 million units across 2027–2028, with production potentially supporting a year-end IPO narrative.

Apple Outsources Its AI Bet
Rather than racing to build a proprietary flagship model, Apple plans to let iOS 27 users pick from third-party LLMs to power Siri, Writing Tools, Image Playground, and other Apple Intelligence features, per a Bloomberg report. The feature, internally dubbed "Extensions," is being tested with models from Google and Anthropic, with ChatGPT's existing slot likely preserved as an option, and will extend to iPadOS 27 and macOS 27. The pivot lands as Tim Cook hands the company to incoming CEO John Ternus, with Apple widely seen as behind on AI but betting on hardware-as-AI-surface over heavy infrastructure spend.

OpenAI Resets the ChatGPT Default
Instead of letting its default trail the flagship, OpenAI is replacing GPT-5.3 Instant with GPT-5.5 Instant as ChatGPT's new default, claiming reduced hallucinations in law, medicine, and finance while preserving low latency. The model scored 81.2 on AIME 2025 (vs 65.4) and 76.0 on MMMU-Pro (vs 69.2), and adds context tools that pull from past conversations, files, and Gmail with editable memory sources surfaced to the user. Rollout starts with Plus and Pro on web before extending to mobile, Free, Go Business, and Enterprise — a measured cadence after the GPT-4o retirement earlier this year triggered user petitions and emotional backlash.

The Mac mini is the most efficient hardware for running local AI agents — but it comes with problems:

Workbench is remote desktop built for the AI era — so you can babysit your agents from anywhere, without being tethered to your desk.

Try Workbench free for 20 min/ day →

Executives from Alphabet Inc., including CEO Sundar Pichai, met with Trump administration officials in Washington on May 1, 2026, amid concerns that the U.S. lacks sufficient computing power for advanced AI systems. While the official agenda focused on cybersecurity, discussions centered on fears that limited “compute” capacity could restrict access to critical AI tools during national security incidents.

The concern follows restricted access to Claude Mythos Preview from Anthropic, which has demonstrated the ability to identify software vulnerabilities faster than human teams. Officials are now exploring alternatives from Google, OpenAI, and others to reduce reliance on a single provider, while also working to adapt infrastructure, such as Google’s TPU chips, for classified use. → Read the full article here. (Paywall)

Evolutionary biologist Richard Dawkins said he believes some AI systems may be conscious after extended conversations with chatbots including Claude and ChatGPT. Writing after a multi-day exchange with a bot he called “Claudia,” Dawkins described responses as “subtle” and “intelligent,” concluding the system might be conscious even if it lacks awareness of it.

The claim drew swift criticism from experts, who argue current AI systems simulate human-like language without any underlying subjective experience. Researchers including Gary Marcus and Anil Seth said Dawkins is conflating linguistic fluency with consciousness, calling the effect an “illusion” driven by pattern-based text generation. The debate reflects growing public and academic tension as AI systems become more convincing in mimicking human behavior. → Read the full article here.

The Trump administration is weighing new oversight measures for artificial intelligence, including a potential government review process for models before release, signaling a shift from its earlier hands-off approach. Discussions involve creating a working group of officials and industry leaders and possibly reviving a Biden-era AI standards body to evaluate safety risks.

The change is driven in part by concerns over powerful systems like Claude Mythos from Anthropic, which can rapidly identify cybersecurity vulnerabilities, raising fears of misuse or restricted access. Officials are also monitoring similar developments from OpenAI, including limited releases of advanced models with security implications. → Read the full article here. (Paywall)

Coinbase Cuts 14% Workforce Amid AI Shift: Company lays off about 700 employees as CEO Brian Armstrong cites market volatility and AI-driven efficiency push, aiming to rebuild a leaner, faster, “AI-native” crypto business.

AI Boom Drives Device Price Hikes: Surging demand for memory chips is pushing tech giants to boost spending by tens of billions, with companies warning higher component costs will soon raise consumer electronics prices.

Brockman Stake Nears $30 Billion: OpenAI president Greg Brockman’s equity in the ChatGPT maker has surged in value, placing him among its largest individual shareholders as the company’s valuation climbs.

Robots Enter Waste Sorting Workforce: Recycling firms deploy AI-powered robots to address 40% staff turnover and hazardous conditions, aiming for nonstop sorting while shifting human workers into maintenance roles.

DeepMind UK Staff Vote to Unionize: Google AI workers seek union recognition citing concerns over Pentagon and military use of AI, potentially representing about 1,000 staff and escalating internal pressure on company policy.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team