← Tue 22 Sept, 07:30 →

☕ Tech digest for September 22, 2026

⚡ TL;DR

  • Grok 4.7 backfire: Elon's new model costs 2x more per token than 4.6 and scores worse on benchmarks—devs are not happy
  • Jev steals the spotlight: TypeSafe's decision model is 238x cheaper than GPT-6 Astra and 100x faster, sparking "is this the LLM killer?" debate
  • MiMo-V2.6 tops the charts: Xiaomi's open model now ranks #1 on Artificial Analysis, native multimodal, 1M context
  • Python on Cloudflare Workers is live: Full Python support (FastAPI, Django, Flask) now generally available
  • The AI narrative is tanking: 60-70% of Americans have negative views—the industry needs a better story

🚀 Big announcements

🤖 Jev launches as a "System One" alternative to frontier models

TypeSafe's new decision model arrived with $40M in funding and claims to be 238x cheaper and 100x faster than GPT-6 Astra. It's optimized for discrete decisions (classification, routing, tool picking) rather than human chat, and can spit out hundreds of parallel outputs with confidence scores for each.

Context

Jev represents a shift in how people think about AI workload. Instead of throwing a frontier model at every problem, you pair a small, fast decision model (Jev) with bigger models only when deep reasoning is needed. It's trained with RLCD (Reinforcement Learning for Calibrated Decisions) rather than traditional LLM pretraining.

x.com ↗

🎯 Xiaomi's MiMo-V2.6 becomes top open model, beats frontier labs

Xiaomi dropped MiMo-V2.6-Pro and MiMo-V2.6-Flash, both native multimodal (text, image, video, audio), 1M context, and the Pro model ranks #1 on Artificial Analysis Intelligence Index at a quarter of frontier lab prices. Landed on the Pareto frontier without the billion-dollar training bill.

Context

MiMo v2.6 followed a massive RL scaling run ($2.6M for Pro, $0.9M for Flash) and shipped with a full technical report. This is the open-source community saying: we can match or beat the labs on intelligence and cost.

mimo.xiaomi.com ↗

🐍 Python Workers go GA on Cloudflare—full frameworks work

FastAPI, Django, Flask—all just work. No surprises. Python is now a first-class language on Cloudflare's edge platform with bindings, database connections, and untrusted code execution all baked in.

x.com ↗

🖥️ Devin adds SSH and CLI for remote agent steering

You can now SSH directly into Devin's VM, poke around the code, forward ports to test locally, or scp files back to your machine. Also launches Devin Cloud sessions from the terminal with /cloud.

x.com ↗

📊 OpenAI's internal math model solved 100+ long-standing proofs

An unnamed model trained from August 28 resolved Navier-Stokes and over 100 other open problems across mathematics. The math advisory board is helping OpenAI figure out how to announce breakthroughs responsibly.

Context

This is a big deal but also shows a governance shift: labs are now thinking about how not to disrupt fields when they solve hard problems.

x.com ↗

🤝 Deel's Akai agent tool added $140M ARR in 90 days

Finance, HR, accounts payable, compliance—Deel built >8,000 agents doing the work of 600 employees, doubled revenue-per-employee, and is now selling it as a product. That's the playbook: automate internally, then productize.

x.com ↗

⚡ Hugging Face Tokenizers v1 (Rust) ships: 30x faster

Multi-language support, thread scaling, tiny package size. Up to 30x faster tokenization, rivaling the best in the ecosystem. Available now.

x.com ↗

🎬 vLLM offloaded video decoding to GPU: 2x throughput

Integrated PyNvVideoCodec to move video codec work from CPU to NVDEC. At 8x H100, this gives 2x throughput for video captioning and metadata extraction at scale.

vllm.ai ↗

🔥 What people are fighting about

💔 Grok 4.7 turns out to be worse (and more expensive) than 4.6

Developers ran benchmarks and found Grok 4.7 costs 2x more per token than 4.6 and scores lower on most benchmarks despite claims of better token efficiency. Theo has posted detailed performance analysis showing regression on Fishslop and other tasks. The response: "How do you ship a model this bad?"

x.com ↗

🎯 Is Jev overhyped or genuinely revolutionary?

Jev's launch sparked fierce debate: some devs see it as the future (smaller, faster, cheaper decisions), others say it's just marketing and only works for non-critical problems. The consensus: great for routing and classification, not a replacement for frontier models.

x.com ↗

🤖 Stop calling Claude your cofounder, it's a tool

Gergely Orosz called out engineers for humanizing AI as a "cofounder" or "collaborator." He argues it's the same category error as saying your keyboard is a cofounder—impressive, useful, but not your peer. Got heated replies about whether AI has crossed some threshold.

x.com ↗

💰 Frontier labs are getting greedier on pricing

OpenAI and Anthropic keep raising prices. Meanwhile, MiMo, Jev, and open models are racing to undercut them. Startups are ditching OpenAI for cheaper alternatives. The narrative: frontier labs are vulnerable if they don't match cost improvements.

x.com ↗

🏗️ AWS made the AI industry soft, and no one is ready for the hard problems

Ethan Mollick's take: AWS handled all infrastructure headaches for so long that everyone forgot how to build. Now with AI infra fragmentation, teams are slapping together "27 model router products" instead of building real inference providers. The gap between labs and everyone else is widening.

x.com ↗

📧 What happens when millions of agents hit email and marketplaces?

Speculation about agents DoS-ing inboxes and marketplaces. Greg Isenberg asked: how do you scale infrastructure when 10M agents start making purchases, and Amazon already blocked Muse? Fascinating governance question no one's answered.

x.com ↗

🧑‍💻 Web dev corner

🐍 Python is now production-ready on the edge with Cloudflare Workers

Python Workers GA means you can deploy FastAPI, Django, or Flask to 300+ edge locations with the same feel as Node. Bindings, databases, untrusted code—all work. Gamechanging for Python devs who wanted edge computing without learning a new runtime.

x.com ↗

⏰ Neon Functions now supports cron triggers and custom domains

Schedule PostgreSQL functions with cron syntax, point custom domains at them. Pairs nicely with scale-to-zero. Makes building serverless workflows on Postgres much less painful.

x.com ↗

🧠 Jared Palmer built Kev: a Jev-like decision model on Qwen3.5

Kev (0.8B, 4B, 9B checkpoints) is an open competitor to Jev, refactored on Qwen and comes with a fine-tuning script. Smaller and easier to run locally. Palmer is planning to scale it up to Qwen3.8-27B.

x.com ↗

🏗️ How should you design codebases for AI agents?

Matt Pocock asked which pre-AI codebase patterns still matter (deep modules, adapters, seams) and why. Lots of discussion about whether tight module boundaries actually help agents or just limit them.

x.com ↗

🤖 AI review bots are finally useful (Cursor Bugbot > Devin > CodeRabbit)

Kent C. Dodds ran the data and found Cursor's Bugbot caught real issues most often over 2 weeks on the Kody project. All three are solid, but rank differently by quality. If you're not using them, you're leaving bugs on the table.

x.com ↗

✨ Disposable Objects API makes test cleanup less painful

Use disposable objects to tie cleanup to setup scope instead of running cleanup after assertions. Tiny change, big impact on memory and flakiness in test suites.

epicweb.dev ↗

🎨 Sarah Edo built a Jev-powered Chrome extension for WebMCP tools

Side panel that auto-picks the right page tool on every keystroke, fills arguments, shows confidence. Tested on a shopping cart—Jev routing to the right tool instantly. Decision models + WebMCP is an elegant combo.

x.com ↗

📬 From the newsletters

📻 AI's real problem: it's losing the narrative to dystopian fears

The Deep View reports 60-70% of Americans hold negative views of AI despite tangible benefits. The industry has tried the "cure diseases, infinite abundance" pitch; no one believes it. What's needed: a Steve Jobs-level storyteller, not more platitudes about AGI.

archive.thedeepview.com ↗

🏢 Salesforce's headless AI strategy: let customers pick their own model

Interview with Shibani Ahuja on how Salesforce scaled agentic AI without losing control. They built a "headless" architecture that lets any model access Salesforce data safely. Also introduced Koa, their CRM reasoning model built on NVIDIA Nemotron.

youtube.com ↗

💼 Data center jobs pay 42% more than equivalent roles

Wall Street Journal reporting data center salaries carry a fat premium. As infrastructure gets critical, human ops matter.

wsj.com ↗

🧪 Research & open source

🎯 LangChain integrated Jev as a production evaluator

Jev-as-a-judge now live in LangSmith. Score every trace instead of sampling, check more criteria per trace without costs climbing. Also released SemIf, an open decision model, free for a week.

langchain.com ↗

📄 LlamaIndex added grounded confidence scores for document extraction

When extracting data from docs, each field now gets a confidence score and a bounding box back to the source. Helps decide what can be auto-approved vs. what needs human review.

developers.llamaindex.ai ↗

🔢 NVFP4 KV Cache: 4-bit KV caching for long context inference

SGLang + NVIDIA + Qwen co-developed NVFP4: 4-bit KV cache that packs ~1.78x more context into GPU memory and speeds long-context decode by up to 78%. Near-lossless accuracy on hard benchmarks.

x.com ↗

🔌 Hermes Agent got an official Claude SDK plugin

Hermes can now use Claude without tradeoffs via the direct SDK. Re-enabled Claude Code subscriptions in Hermes.

x.com ↗

🤖 Papers on agent ontologies, skill evolution, and retrieval strategies

Research dropping on self-evolving ontologies for data agents (26.7-point gains on benchmarks), SkillLift (skill improvement at 40-70% less token cost), and Question's Gambit (better retrieval as the opening move for research agents). All practical patterns for building agent systems.

x.com ↗

🤖 Eidon AI open-sourced 1,274 hours of egocentric robotics data

Company winding down but left a gift: 13,451 videos of humans doing everyday tasks (laundry, cleaning, cooking) with paired 7-IMU arm tracking. 9TB, CC-BY-4.0. Open source means the work outlives the organization.

huggingface.co ↗

🍿 Worth a read

📝 Simon Willison on Jev and the "System One" category of models

Excellent explainer on why decision models are different and why they matter. Jev isn't trying to be a better LLM; it's a different tool built for a different job (fast, cheap decisions with confidence scores).

simonwillison.net ↗

✍️ "Delegate coding, never delegate understanding"

Francois Chollet on the line between what you can offload to AI and what you have to keep in your head. If code was your source of truth for understanding, you need a new artifact now that code-generation is cheap.

x.com ↗

⚙️ What Sun Microsystems got wrong (and what we can learn)

Deep technical post on infrastructure, interfaces, and why Sun lost despite being ahead on so many fronts. Lessons for building AI platforms that don't fragment like Unix did.

bcantrill.dtrace.org ↗

🎯 OpenCode 2.0: agents that can modify themselves in real-time

Kit at Anoma finished a blog post on OpenCode's ability to rewrite its own harness during a session. Agents bootstrapping themselves. Rad and a bit scary.

anoma.ly ↗