San Jose Tops US Cities Most Exposed to AI
ALSO: Chinese hackers wield DeepSeek, Thomson’s $40M AI
Welcome to Guru99 AI Report!
Top Story: Where you live may decide how much AI reshapes your work. We’re also unpacking a $40M homegrown legal model, an anonymous coding tool developers can’t stop using, and why cheap open-weight AI just became a security headache. Let’s dig in.
🗺️ Where You Live Shapes How Much AI Changes Work
Brief Buzz:
The Hiring Lab at Indeed scored 386 US metro areas on their exposure to AI — how much of the local workforce generative AI could reshape. Tech-heavy metros land at the top and hands-on economies at the bottom, but exposure doesn’t mean job loss.
- San Jose takes the No. 1 spot in the new metro-level exposure ranking with 59, followed by Seattle (57), Washington, D.C. (54), San Francisco (53), and Austin (52).
- The average across 386 metros is 44, on a scale running roughly 40 to 60.
- Two defense hubs crack the top 10 — Lexington Park, Maryland and Huntsville, Alabama — both engineering-heavy.
- The lowest scores go to metros built on manufacturing, healthcare, and retail work.
- 55% of US adults under 30 are now more concerned than excited about AI.
💡 Why Should You Care?
High exposure means your tasks might change, not that they’ll disappear. Indeed’s own qualification matters here: it’s up to local employers whether those changes turn into layoffs or into better tools.
🔓 Chinese Hackers Doubled Attacks Using Cheap Open-Weight AI
Brief Buzz:
Taiwanese threat-intelligence firm TeamT5 told Bloomberg that Chinese state-linked hacking groups have more than doubled their attacks since folding open-source AI into their operations. Their tool of choice isn’t a cutting-edge model but DeepSeek — cheap, customizable, and only lightly protected.
- Grimfengxi wrote exploit code with DeepSeek, while Teleboyi pulled 1,000 IP addresses and mapped the target’s domains.
- A third group, Huapi, hit a Taiwanese company’s email system using a model the researchers believe to be DeepSeek.
- Cost is the main driver: DeepSeek V4 runs about $0.87 per million output tokens, while Moonshot’s Kimi K3 costs $15.
- TeamT5 saw no incidents with Kimi K3 — capable and open-weight, but too expensive to run at scale. Yahoo!
- Western models still surface too — one group ran Claude Code through a Taiwanese network by posing as authorized testers.
💡 Why Should You Care?
Baked-in guardrails offer real protection; the kind you can download and strip out do not. Although AISI is watching frontier models get sharper, the low-end may in fact be the harder problem to deal with.
🥷 Nobody Knows Who Built This Free Coding Model
Brief Buzz:
Ox Alpha appeared on OpenRouter on August 20 with no company name and no price attached, and its ability to process a million tokens at once sent developers flocking. Five days on, no one has claimed it, and the evidence points to Z.ai in China.
- For now, 1,048,576 tokens of context, 131,072 tokens of maximum output, plus text, image, and video input are all free.
- The viral 80% DeepSWE figure came from just 10 of 113 tasks; further community runs land near 63%.
- Ox Alpha isn’t listed on DeepSWE’s official board, where Fable 5 sits at 70%.
- OpenCode claims a capacity of 100 trillion tokens per day — a figure that’s unverified and has drawn open skepticism.
- Patrick Collison, whose Stripe is acquiring OpenRouter, called it “very impressive.”
💡 Why Should You Care?
Coding help just became free for a week. The drawback is that an anonymous provider keeps your prompts — fine for side projects, but not for proprietary code.
⚖️ Thomson Reuters Built Its Own AI for $40M
Brief Buzz:
Thomson Reuters has released Thomson, its first in-house AI model. It’s built on Alibaba’s open-source Qwen and trained on decades of Westlaw and Reuters material. The aim: stop renting AI from Anthropic and OpenAI and start owning it.
- The bill came to $40 million over two years for staff and compute. The last training run cost roughly $450,000.
- By its own standards, Thomson matches Claude Opus 4.8 and outperforms GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro.
- An independent review found gaps: Thomson stumbled on LegalBench, reasoning, and coding, and needed extra test-time compute.
- Its first job was document review at CoCounsel Legal — a product that still mainly runs on Claude.
- A smaller open-weight version is headed to Hugging Face, but only for academic institutions and non-commercial use.
💡 Why Should You Care?
Lower-cost legal AI could translate into cheaper legal services — and if a $40 million model can match ones that cost a billion dollars, other data-rich firms may follow the same path.
🏭 Oumi Lets Enterprises Own Their AI, Not Rent It
Brief Buzz:
Seattle-based startup Oumi wants companies to own their AI rather than rent it from big labs. On August 11 it launched a Compounding AI Factory that builds custom models from a prompt, deploys them in one click, and keeps improving them on your own data.
- One click deploys GPUs that scale automatically with traffic and drop to zero when idle, cutting wasted compute costs.
- The models retrain automatically on live production data, turning real-world failures into fresh training signal.
- A public command-line tool built for agents lets coding agents run training and evaluation just like a developer would.
- Everything — the model weights, training data, and build recipe — is exportable under the Apache 2.0 license.
- In Oumi’s own case study, a top-five US bank used it to help modernize 100 million lines of legacy code.
💡 Why Should You Care?
For businesses the real issue is control: cheaper, more specialized models let sensitive data stay on-premises and improve over time. The drawback is that you still need clean, well-organized data for any of it to work.
🤖 Meta’s New Open AI Runs Right On Your Device
Brief Buzz:
Meta is back in the open-model race. Its new Muse Glimmer is an agentic model with 30 billion parameters that runs on a single consumer-grade GPU — no cloud required. The launch arrived with a manifesto from Mark Zuckerberg arguing that AI should empower everyone, not just a handful of labs.
- It has 30 billion parameters yet runs locally on one consumer-grade GPU under the permissive Apache 2.0 license.
- It’s Meta’s first fully open release since it discontinued its Llama line in April and launched the proprietary Muse Spark.
- On Meta’s own benchmarks it beats same-size rivals at agentic tool use, but it trails Qwen on computer use and terminal tasks.
- Zuckerberg’s essay, “The Future is for Everyone,” argues the biggest danger is concentrating superintelligence in a few institutions.
- Meta also plans to release the open-source version of its Muse Spark 1.2 in the coming weeks, deepening its open-source push.
💡 Why Should You Care?
Advanced AI agents could run privately on your own device — no subscription, and no data sent to the cloud. Just weigh Meta’s “AI for everyone” pitch against how much AI power it’s amassing.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
