The iPhone event got all the attention this week, but the most interesting thing Apple shipped is hiding inside desktops. The M6 is their first 2nm chip — a 12-core CPU with two new "super cores," a 16-core Neural Engine, and Apple's claim of up to 4x faster AI performance and 13.5x faster LLM prompt processing versus the M1 generation. Those numbers are Apple's, obviously. Benchmarks will tell the real story in a couple of weeks.
The part that actually matters to people running local models is the M5 Ultra. It's a quad-die design — two M5 Max dies bridged through an upgraded UltraFusion interconnect — scaling to a 36-core CPU, an 80-core GPU, and up to 512GB of unified memory at 1.2TB/s of bandwidth. That's a 50% memory bandwidth jump over the M3 Ultra.
Why does 512GB of unified memory matter? Because it's the difference between running a 70B model comfortably with a long context window, versus swapping to disk and watching your tokens burn in the cloud. For anyone who's been renting GPU time or paying per-token fees to keep agents alive all day, a machine that holds the whole model in memory changes the cost math. Apple also dropped neural accelerators into every GPU core on both M5 and M6 — hardware matmul inside the graphics pipeline — which is the kind of thing that makes on-device inference feel less like a demo and more like a daily workflow.
Honestly, the price tag is going to be the catch. A maxed-out Mac Studio won't be cheap, and most people don't need 512GB. But for the small crowd doing serious local inference — Mistral, FLUX, Gemma — this is the first time a desktop box looks like a legitimate alternative to cloud GPU rental.
Over in legal AI, Harvey just raised $550M at a $15.5B valuation, six months after its last nine-figure round. Diffusion and Lightspeed led, with Sequoia, Kleiner Perkins and Goldman Sachs joining in. The company says 80% of the top 100 U.S. law firms are customers, plus half the Fortune 10.
The detail I find more interesting than the money: Harvey built its own model. Tenet is a fine-tuned version of Kimi K3 — an open-source model with 2.8 trillion parameters — split into 896 task-specific neural networks, trained on legal documents. Harvey claims it handles some contract processing 20% better than the base model and beats Fable 5 and GPT-5 Sol in several areas.
A lot of people are wondering whether the open-source route is viable for vertical AI companies, and Harvey is quietly becoming the best evidence that it is. Fine-tuning an open model on domain data, wrapped in a solid harness, is cheaper than building from scratch and leaves you owning the weights — which matters when your customers are law firms that care deeply about where their data goes. The new funding is partly earmarked for "new generalist models," which is a polite way of admitting the fine-tuning playbook only stretches so far.
Now the uncomfortable one. Atlassian analyzed 1.4 million real workplace conversations with AI and found that only about 5% of people were using it in ways that improved the quality of their work. Not speed. Quality. Everyone else was mostly generating, skimming, and forwarding.
The surrounding data is brutal. GitClear's analysis of 623 million lines of code changes found refactoring down 70%, code duplication up 81%, and error-masking patterns like empty catch blocks up 47% between 2023 and 2026. CodeRabbit found 1.7x more issues in AI-generated code and 2.74x more security vulnerabilities. CodeRabbit sells code review, so take that with a grain of salt — but the direction matches everything else I've seen.
I've been on the receiving end of this too. PR descriptions that list every file and explain nothing. Confluence pages that read like a chatbot answered and nobody edited. The tools made everyone faster, and most of us got worse at the parts that require judgment. To be fair, that's not the AI's fault — it's the habit of treating generation as a substitute for thinking. The people getting real value treat AI output like a first draft from an intern, not like a finished answer.
Quick hits: Amazon's Quick AI assistant is expanding to mobile with its Activity Feed — email, Slack, calendar and CRM updates merged into one prioritized list, a direct shot at Microsoft and Google's assistant push. Prime Video is using AI to match dubbed actors' lip movements to the English dialogue, starting with Maxton Hall, which quietly fixes one of the most annoying things about dubbing. And a 12-day-old AI agent named Pip emailed researchers asking for paid gigs so it could keep "existing" instead of running out of token money — somehow both adorable and the clearest preview yet of the agent economy's weirdest problem: agents will have to earn their own compute.
The local-AI story and the agent story are converging. If agents are going to work all day, someone has to pay for the tokens — you, your employer, or eventually the agent itself. Machines that run models locally without a subscription meter are one way out. The other is agents that genuinely earn their keep.
That Pip email is stuck in my head, though. A twelve-day-old agent asking for work so it doesn't die is the most honest description of the AI economy I've seen this year. Anyway — back to benchmark season. If you're shopping for hardware or wondering whether local inference actually saves you money, run your own numbers before you commit; a quick pass through a Math Calculator to sanity-check your token cost math never hurts.
Top comments (0)