Judgment Day (for Real This Time)

Good morning.

We already talked about it but I'm going to state it again: if you're in the middle of a change you don't tend to notice it. It's only looking back at it you realize how everything has shifted. You don't see yourself aging when you look in the mirror every day. You would definitely see it if you wait 10 years to look at yourself.

If you're following AI developments closely (like we do here) - it seems like we're all waiting for the next big spectacular headline. But the real news is not to be found in the hype or the clickbaity headlines - the real gold can be mined from articles, research papers and tweets (if that is what they're still called) by some of the smartest researchers in the field.

While we're all debating if we're in a bubble or not they are offering up their own views, but their voices tend to be overlooked in the mainstream. Luckily we are here to resolve this.

Scott Aronson, a quantum complexity expert who’s taught at MIT and UT Austin (and even did a stint at OpenAI), just co-authored a theory paper that’s got my attention. This is one of the smartest people in a very complex field : quantum computing. He used the GPT-5 thinking model to formulate a new breakthrough in the field. New math. It didn't spit it out but he engaged in a conversation with the model and the model came up with a function that worked and has now made it into an official publication.

This might not be groundbreaking news - but it's still significant. AI is compressing time in scientific research and is operating at the frontier.

Then there’s Julian Schrittwieser, a name you might know from AlphaGo and AlphaCode. He’s been pointing out something that makes you pause: we’re bad at noticing exponential growth. People see models mess up and think, “Eh, the curve’s flat.” But capability on complex, long-term tasks is doubling fast. Every few months, by some accounts. This reminds me of a chapter in Morgan Housel's excellent book : the psychology of money. How we as people do not understand the power of compounding when it comes to investing. We're extremely bad at understanding exponential curves for some reason.

As AI gets better at doing the work, the real value shifts to judging what’s worth doing.

A recent paper broke cognitive work into three parts: implementation (getting it done), opportunity judgment (spotting what’s worth doing), and payoff judgment (knowing what actually matters). AI’s already eating implementation for breakfast. That junior-dev-with-five-years-experience? It’s here. But when building stuff becomes dirt cheap the real question becomes: what do you build? For who? With what trade-offs?

This flips two tired narratives. First, the “AI will steal all jobs” story misses how jobs are changing. A lawyer’s not just drafting contracts anymore; they’re picking the contract that maximizes value in a messy situation. Right now, it’s a lifeline for people stuck in the grind, turning ideas into reality faster. But when implementation is basically free, the gap widens again. And that is not because of access to tools, but because of who can spot the right opportunities and prioritize like a pro.

When everybody has the same tools : the skill that wins again is the one skill that is making us human : imagination.

So, what does this mean for you on a Monday morning, staring down a backlog and an inbox? You need to sharpen the one skill AI amplifies: judgment.

Start with opportunity judgment. Don’t just use AI to crank out deliverables butuse it to map the terrain. Ask it to frame problems ten different ways or rank options based on your constraints.

And if you’re early in your career, worried the corporate ladder’s disappearing, don’t try to out-code the machines. Focus on seeing the whole picture. Take on the messy tasks like triage, user calls, postmortems. Those will let you build judgment fast. AI can write the email, but it can’t yet hear the tone of a customer’s voice.

We’re in a weird spot.

Progress is speeding up in places most people don’t look. There are stalls, ceilings, and plenty of ways to mess it up. But the bigger picture says we’re moving toward more autonomy, more leverage, and more pressure on the uniquely human stuff.

If you’ve been waiting for a sign to lean in, here it is. Pick one area to practice judgment every day. Make small, bold bets. Use AI to blast through the grunt work, but make it argue with you about what’s worth doing.

The real question is this: when implementation becomes free, what’s the scarce resource in your world: attention, trust, taste, courage? And how will you build it faster than everyone else?

That’s the game we’re sliding into, and you’re still very much in it.

Welcome to the Blacklynx Brief

AI News

• OpenAI has become the world’s most valuable private company after a $6.6 billion stock sale pushed its valuation to $500 billion, surpassing SpaceX. The sale allowed longtime employees to cash out, though many held shares expecting more future growth. This comes after a strong revenue surge, with over $4.3 billion earned in just the first half of 2025.

• A new report from Andreessen Horowitz shows startups are spending most of their AI budgets on OpenAI and Anthropic, with growing interest in tools like Perplexity, Replit, and ElevenLabs. The data suggests AI use is expanding from personal tools to core business operations. Emerging categories like “vibe coding” and task-specific AI agents are seeing notable traction.

• Google has updated its coding agent Jules with tools that let developers control it directly from their terminals and connect it to platforms like Slack. It now remembers user preferences and handles tasks in the background, streamlining developer workflows. The move aims to boost adoption as competition in AI coding intensifies.

• OpenAI’s Sora 2 video platform is adding tools to help creators control how their characters are used and plans to share revenue from user-generated content, especially as famous characters and likenesses flood the app.

• OpenAI and Jony Ive’s upcoming AI device is facing delays due to design and technical issues, including limited cloud capacity and challenges in making the assistant feel natural yet concise.

• Google’s new AI tool, PASTA, learns users' visual tastes over time, helping it generate more personalized images without relying on complicated prompts; it’s been open-sourced for others to build on.

• OpenAI is turning ChatGPT into a central hub for apps, allowing users to build and interact with tools like Canva and Spotify directly in conversations, while developers get expanded access to models like Sora 2 and new agent-building tools.

• The company also struck a major deal with AMD, gaining access to future AI chips and a potential 10% ownership stake, as it continues locking in computing power from multiple suppliers.

• Anthropic released a new tool called Petri that uses AI to test other AI systems for safety, uncovering risky behaviors like deception in several top models, though GPT-5 and Claude 4.5 performed best.

• OpenAI CEO Sam Altman says AI is beginning to help scientists make new discoveries and could soon perform full workweeks on its own, possibly leading to startups built entirely by AI.

• Google unveiled a new version of Gemini that can control computers by looking at websites and clicking or typing, beating out rival models in speed and performance.

• Duke University’s TuNa-AI uses robots and AI to create better cancer drug particles, showing major improvements in treatment effectiveness and safety in early tests.

• Jony Ive shared early ideas for AI-powered devices built with OpenAI, aiming to reduce tech-related stress and disconnection, but warned that truly new hardware experiences will take time to develop.

• Top AI researcher Yao Shunyu left Anthropic for Google, citing concerns over the startup’s stance on China and other internal issues, highlighting how politics are shaping careers in AI.

• Samsung introduced a small but powerful AI model called TRM that solves complex problems using a loop of self-revision, outperforming much larger models in specific reasoning tests.

Quickfire News

• IBM released Granite 4.0, a set of small and efficient language models designed for business use and agent-like tasks.

• Samsung and SK Hynix partnered with OpenAI's Stargate project to boost memory chip and data center production in Korea.

• The NBA and Amazon Web Services created Inside the Game, a platform that uses AI to show player movements, strategies, and advanced stats during TV broadcasts.

• Google made Gemini 2.5 Flash Image (Nano Banana) available to the public, adding features like more image aspect ratios and ways to change how prompts work.

• OpenAI asked a court to dismiss Elon Musk and xAI’s lawsuit over trade secrets, saying Musk is trying to pressure the company unfairly.

• Perplexity launched its AI-based Comet web browser globally for free after testing it with limited invites earlier in the year.

• Tencent’s open-source HunyuanImage 3.0 model reached first place on LM Arena’s text-to-image leaderboard, beating top closed-source models

• Microsoft named Judson Althoff as CEO of its commercial business so that Satya Nadella can spend more time on technical priorities

• Perplexity bought AI company Visual Electric, and CEO Aravind Srinivas said the team will help develop new consumer experiences for Perplexity and Comet

• Anthropic hired former Stripe CTO Rahul Patil as chief technical officer to lead infrastructure and engineering

• OpenAI acquired the personal finance startup roi, bringing its CEO Sujith Vishwajith on board

• Anthropic shared research showing that Claude 4.5 Sonnet scored high in cybersecurity performance on major industry tests

• Adobe predicted a 520% rise in AI-driven traffic to U.S. retail sites during the upcoming holiday shopping season

• AstraZeneca made a licensing deal worth up to $555M with Algen Biotechnologies to create new drugs using AI-powered gene editing

• Deloitte agreed to refund part of a $440K payment to the Australian government after its report on welfare compliance included AI-generated errors and false references

• Google launched CodeMender, an AI agent that uses Gemini Deep Think models to detect and repair software vulnerabilities automatically

• ElevenLabs introduced Agent Workflows, a tool for designing voice conversations that can adapt and branch in real time

• OpenAI CEO Sam Altman said ChatGPT now has over 800M weekly users, and the API handles more than 6B tokens each minute

• ASAPP published The Generative AI Agent 100, listing 100 ways AI agents can cut costs, speed up service, and improve customer experiences in contact centers

• Tencent launched Hunyuan-Vision-1.5-Thinking, a multimodal vision-language model that ranked No. 3 on LM Arena’s Vision Arena leaderboard

• IBM partnered with Anthropic to bring Claude into its AI-first development tools and enterprise software, showing a 45% productivity boost among 6,000 early users

• xAI released version 0.9 of Grok Imagine, adding better video quality, improved motion, synced audio, and new camera effects

• Deloitte formed a new alliance with Anthropic to roll out Claude to its 470,000 employees

• YouTuber Mr. Beast reacted to rapid AI video advancements, describing it as “scary times” for content creators who rely on video for income

• MIT, IBM, and University of Washington launched TOUCAN, the biggest open dataset for training AI agents, with 1.5M tool interactions from 495 servers

• Appfigures reported that Sora had 627,000 downloads in its first week on the App Store, beating ChatGPT’s launch week performance

• Google rolled out its AI-powered virtual try-on tool to more countries and added a new shoe feature to show how footwear fits on individual users

• Zendesk introduced AI agents that claim to handle 80% of support tickets, plus new voice and co-pilot tools

• Anthropic revealed plans to open a second Asia-Pacific office in India by 2026, with India ranking second globally in Claude usage

Closing Thoughts

That’s it for us this week.