GPT 5.6 (Highthink) is out here doing Claude's lazy bit now. We really crossed a threshold. Anyone else see this?

8/8 The price is the story. When a category leader with $500M in ARR trades at 3x, the moat is melting. Bending Spoons is betting the melt is slow. Every AI-native builder is betting the opposite.

7/8 The real irony is that Airtable sat on the perfect AI product. Turning a prompt into a working app inside Airtable was the obvious paid feature, and they never shipped it. A distinct lack of imagination.

6/8 The precedent is Meetup. Subscriptions jumped to €29 a month after the acquisition with little added to the product, and users drifted to Luma. Airtable users have been complaining about pricing and limits for years.

5/8 Anyone can vibe-code a working Airtable clone in a weekend, so the moat was always inertia and data gravity. Inertia keeps paying the bills for a few more years. Data gravity is what Bending Spoons is actually buying.

4/8 The timing is the interesting part. Enterprise CRUD and Access-replacement workflows, a big slice of Airtable's revenue, are low-hanging fruit for AI agents, and renewal pressure in that category has been visible all year.

3/8 Bending Spoons' playbook is stated openly in its IPO prospectus: buy sticky products, fold in shared services, cut staff, maximize monetization. AOL, Eventbrite, Vimeo, Evernote, WeTransfer, Komoot. Now Airtable.

2/8 The math is brutal. about $500M of ARR sold at roughly 3x after raising $1.4B. The deal is cash-free and debt-free, so Airtable's net cash implies an equity value closer to $2.25B.

Bending Spoons is buying Airtable for $1.285B in cash, its first acquisition since going public on Nasdaq last month. The company peaked at an $11B valuation in 2021 and is now selling for roughly what it raised.

1/8 Bending Spoons is buying Airtable for $1.285B in cash, its first acquisition since going public on Nasdaq last month. The company peaked at an $11B valuation in 2021 and is now selling for roughly what it raised.

Pricing runs $2 per million input tokens, $6 per million output. Open weights arrive next week. The gap between what closed labs charge and what open weights give away keeps shrinking. The moat is quietly moving from owning the model to owning the work around it.

The range keeps expanding: a full silicon flow, cutting an accelerator from 8,298 gates to 678 and shrinking die area 81 percent. A 365-day e-commerce simulation where it turned 100k yuan of capital into 416k, finishing 38 percent ahead of the runner-up.

The training recipe is the underrated part. Realistic work environments scaled on task, workspace, and harness. One universal reward system instead of task-specific verifiers. An online data balancer keeping batches stable. They are scaling work competence, not just test scores.

They also entered it in a live 24-hour contest against 526 human teams. It read the rules, fine-tuned five models, fused them with weighted voting, and made 45 submissions. Final accuracy 0.853. It beat 87 percent of the human field.

Then they handed it a research paper and said: reproduce this. In about five days it wrote 7,600 lines of training code and ran 33 rounds of GPU experiments. Then it went further than the paper, testing 18 of its own ideas and gaining another 2.7 points on AIME24.

The demos matter more than the scores. Qwen ran the model for 10 days straight building a repo that improves itself. 265 commits, 127 PRs, 151 issues, all claimed, coded, tested, and merged by agents. No human in the loop.

The benchmarks sit right next to the closed frontier. PaperBench 93. TerminalBench 2.1 86.6. OSWorld-Verified 86.1. Roughly level with the best of Claude and GPT on many, behind on a few. No open-weight model has ever been this close.

Alibaba officially released Qwen 3.8-Max. 2.4 trillion parameters, 95 billion active. And a first: a Qwen-Max-class model is going open-weight. Weights land on Hugging Face next week. The preview has been behind their paywall since July. Now anyone can take the model home.

9/9 The durable value in AI infrastructure is consolidating around physical capacity plus the software that can co-design with it. Pure software orchestration does not stay independent for long. It gets bought by whoever owns the racks. Nscale just proved it again.

8/9 But the execution bar is high. Nscale is energizing data centers, expanding to the US, servicing a 200k chip Microsoft deal, and now integrating a 7-year-old software company. All at once, on borrowed credibility and a lot of borrowed capital.

7/9 The deeper lesson: optimizing one layer at a time is over. The winners co-design the whole stack, software and hardware together. That is why hyperscalers build their own chips and frameworks. Nscale is trying to run the same playbook at two years old.

6/9 The real risk is gravity. The license is fine. Even with neutral governance, Anyscale's engineers now optimize Ray for Nscale racks first. That helps Nscale customers. For everyone else it is a slow drift to watch. Privileged substrates tend to become default ones.

5/9 The interesting part is open source. Ray is Apache 2.0, governed by the PyTorch Foundation, with contributors from Google, NVIDIA, Microsoft, Alibaba. Nscale is joining the Foundation as a platinum member. The promise is that Ray stays open and portable across clouds.

4/9 This is the pattern now. Groq to NVIDIA. Aleph Alpha to Cohere. Silo AI to AMD. Now Anyscale to Nscale. Independent AI software layers are acquisition targets. The capital and the demand sit on the physical infrastructure side, so that is where the consolidation starts.

3/9 For Anyscale this is a solid outcome, and it caps the upside early. Roughly $440M raised, a $1B Series C in 2023, now $1.65B out. The standalone Ray business never compounded into a $5B independent. The infrastructure gravity well won.

2/9 The logic is vertical integration. Nscale owns the land, power, data centers, and GPUs. Anyscale owns the software that schedules work across all of it. Buy the layer that sits on top of your racks and you stop selling GPU hours. You sell a platform.

1/9 Nscale is buying Anyscale for $1.65B. Nscale is a 2-year-old UK neocloud, $14.6B valuation, building GPU data centers at multi-gigawatt scale. Anyscale is the company behind Ray, the open source framework running a huge share of AI training and serving.

3/3 Open-source agents are eating the AI assistant category. OpenWorker, opencode, aider, open interpreter. All local, all community-driven. The moat in this space was always the workflow. Models are commodities now. Workflows want to be open.

2/3 The architecture is the real story. Local-first, model-agnostic, approval-gated actions. 25+ connectors, MCP support, any provider or fully local via Ollama. Everything stays on your machine. The open agent thesis is shipping.

1/3 Andrew Ng just open-sourced an AI coworker called OpenWorker. Runs on your desktop, plugs into Slack and GitHub and Gmail, hands you finished deliverables instead of chat. 10k stars in days.