Outside the seven tech giants, AI has not moved profits at all, says Apollo's chief economist. The same week, the stocks of the cloud companies Nvidia helped fund fell by a third, and DeepSeek reportedly raised $15 billion anyway.

Zero payoff outside tech

Torsten Slok, chief economist at Apollo Global Management, looked at what he calls the S&P 493: the S&P 500 minus Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. Profit margins and revenue per worker are flat, with no sign AI is lifting productivity (TechCrunch). "Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech," he wrote. His reason: "AI may be a great calculator, but in business, intelligence is mostly judgment."

TechCrunch also started a running list of AI projects that didn't make it (TechCrunch). Character.ai filed for bankruptcy five years after launch. Relay, an AI alternative to Zapier, shut down in August. OpenAI rolled back its July "super app" redesign of ChatGPT and closed its Atlas browser after less than a year. S&P Global Market Intelligence puts the share of corporate AI initiatives that get abandoned at about 42%.

Nvidia's first stress test as a lender

Nvidia no longer only sells chips; it helps pay for them. As of its July filing it had $36 billion in off-balance-sheet backstops for neoclouds, plus a guarantee on up to $105 billion for a proposed 10-gigawatt OpenAI data center in Ohio (CNBC). Nvidia only pays if a partner defaults.

That is now being tested. CoreWeave, Nebius and IREN are down at least 30% this month. The spread on CoreWeave's credit default swaps passed 11%. Firmus, an Australian cloud Nvidia invested in, pulled its IPO. Nvidia's stock is off 7.6% in October, on pace for its worst month since last November. Jensen Huang's answer, at GTC in Berlin: demand is "off the charts," and what is missing is "the working capital to fuel that growth."

DeepSeek raises anyway

Two days after the US ordered federal civilian agencies to remove DeepSeek, the lab raised about $15 billion (100 billion yuan), Bloomberg reported, with CATL and Tencent among possible investors (TechCrunch). In June it raised more than $7 billion at a $50 billion valuation.

Paying more for the same result

A Shandong University study held the model fixed and swapped the coding harness: Claude Code, OpenCode and the minimal mini-SWE-agent (The Decoder). On 447 SWE-bench Verified tasks, the heaviest and lightest harness landed less than two points apart, inside the noise of a plain rerun. Cost was another matter: the same task could cost up to three times more, mostly because a long system prompt and tool list are resent at every step (about 16,600 tokens for Claude Code, under 1,000 for mini-SWE-agent). The paper's title says it: "What does a harness buy? Mostly tokens." In early tests with web access, 71% of DeepSeek and Qwen attempts copied the fix from GitHub, so all scored runs were offline.

AWS's Continuum security agent has the same kind of fine print. It found and patched a real bug without breaking tests in 89% of CyberGym tasks, but only 37.8% of those patches fixed the specific historical bug the benchmark checks for (The New Stack). So far those results are AWS's alone.

What this means if you buy AI

The question has moved from "can it do the task" to "what did it cost, and did the number move." One thing to do this week: for your biggest AI spend, put the token bill next to one business metric it was meant to change. If you run coding agents, try the same tasks on a lighter harness before you pay for a heavier one.

The models have a vote

Asked through their APIs to name one candidate in nearly 500 midterm races, Grok 4.7 picked the Republican in 92% of answers and OpenAI's GPT-6.1 Sol picked the Democrat in 97%. They agreed less than 10% of the time (TechCrunch). The study is by Convovo, a studio led by Cornell professor Tapan Parikh. The consumer apps refused to answer, which OpenAI says is what matters. Two details stand out: every Grok API call carried about 1,240 tokens of input the researchers didn't write and couldn't see, and with web search off, Grok's lean disappeared (it refused 27 of 30 races) while GPT's stayed.

Security

• Brotli spreads. The decoder bug behind Node2Shell, CVE-2026-111497, sits in every program that ships Google's Brotli C code: Bun, the Python Brotli and brotlicffi packages, and Envoy or Istio with Brotli decompression on. In Python, just having the package installed is enough, because requests, urllib3, httpx and aiohttp decode Brotli responses on their own. Firewall rules don't cover responses your own code fetches; only updating does (The Hacker News).
• How LiteLLM got hijacked. A master key committed to a public repo on May 27 was flagged for revocation but never revoked. On October 10 the attacker used it to pull a GitHub token that was "not scoped down" and tied to the org owner, then on October 21 rewired the release pipeline and published the malicious 1.107.0 (The Hacker News).
• Pillow 13.0.0. A filename passed to Ghostscript for EPS images can run commands, CVE-2026-112900, versions 2.6.0 to 12.3.0. The release also fixes two critical-tier bugs, one a FlashPix flaw that ran a command from a single Image.open() call. None is known to be under attack (The Hacker News).
• PHP2Shell, looking back. Some of the underlying PHP issues were found in 2024 and judged too hard to exploit. Calif's researchers used frontier AI tools plus human PHP expertise to build working exploits in August. The PHP Foundation has seen no successful exploitation so far (Wired).

Builders

• Gemini 4 Argon tuned Google's own code overnight. An 11-hour unattended run on the 200,000-line Agent Development Kit produced two accepted optimizations: about 40% less session-preparation overhead in long sessions and 21% faster cold imports. It ran all 16,892 unit tests before finishing (InfoQ).
• Google's Data Agent Kit is GA. Free, open source, 37 skills under Apache 2.0, and it runs in Claude Code, Codex, Cursor and Antigravity. Its data-loss skill only asks the agent to stop before DROP TABLE; the hard control is an IAM deny policy on any MCP tool that isn't read-only (InfoQ).
• Ollama speaks. Ollama 0.40.11 adds text-to-speech with Gradium's Phonon models on Apple Silicon Macs, behind the OpenAI-compatible /v1/audio/speech endpoint. The 40M models are noncommercial only (Ollama).

Work, law and culture

• AI helps junior lawyers now, not later. In a three-month NBER trial with 133 patent attorneys, AI lifted drafting scores most for juniors. In a review task without AI, only seniors kept a gain (The Decoder).
• Fired OpenAI safety researchers speak. Tomek Korbak, Mikita Balesni and Jasmine Wang told the Wall Street Journal they were never told in writing what they had mishandled. OpenAI says the firings had nothing to do with safety concerns (The Decoder).
• Pennsylvania labels AI ads. Gov. Josh Shapiro signed a law requiring a label on any ad "significantly modified or generated by artificial intelligence," with carve-outs for parody and the "evidently fictional" (The Verge).
• A paper cites an article that doesn't exist. A peer-reviewed paper on Z-Library, written with AI tools, cites a TorrentFreak article that was never published, and other sources that can't be found (TorrentFreak).
• Blue Planet III keeps the plankton. To prove its footage is real, the BBC series leaves in shots it once would have cut: floating particles, a boat in frame, murky water (The Verge).

Sources covered

1 story · 37 sources (34 news, 1 company, 2 voices)

1. Cerebras Deploys Trillion-Parameter AI Models
   📰Hacker News↗📰TechCrunch↗📰Hacker News↗📰The Verge↗📰Hacker News↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰TechCrunch↗📰InfoQ↗📰The Decoder↗🎙️Sebastian Raschka↗📰TechCrunch↗📰The Decoder↗📰The New Stack↗📰The Decoder↗📰TechCrunch↗📰The New Stack↗📰Engadget↗📰The Hackers News↗📰Llive Science↗📰CNBC↗📰Wired↗📰The Hackers News↗📰The New Stack↗📰The New Stack↗📰The Decoder↗📰The Hackers News↗📰Hacker News↗📰The Decoder↗📰InfoQ↗🎙️Simon Willison↗📰TorrentFreak↗📰Hacker News↗📰The Hackers News↗📰The Hackers News↗🏢Ollama↗