AI the News That's Fit to Prompt

Apollo’s chief economist says “the marginal cost of producing compute exceeds the marginal benefit it delivers to users,” and five small Qwen3.8 models orchestrated by a shared notebook just scored 93.0% on hard coding problems against Fable 5’s 90.4% at 9% of the cost. Put those next to a $10 billion text-message agent, a three-second voice clone, and a graveyard of products their own platforms swallowed, and today’s stories share one question: who actually captures the value when intelligence gets cheap?

Today’s Headlines

The Price of Compute Under Scrutiny

• The “Core Economic Flaw” – Apollo chief economist Torsten Slok argues that data-center demand depends on the price users will pay, not on power or physical capacity, and that firms will only pay if it lifts revenue, profit, or savings. He cites a recent slowdown in token spending growth and says the demand curve for compute is flat and highly sensitive to price. On Thursday he added a “circularity” warning: 80% of the top 1% of AI spenders are VC-backed. His own concession is the most telling line: “Maybe that will come in the future. Maybe not.”
• GVS5H – The paper’s method has fresh instances of one model decompose a problem and coordinate through a shared filesystem holding a plan, notes, and the current solution. Across eleven models on the 100 latest hard LiveCodeBench problems it produced gains of up to 25.6 percentage points. Orchestrated Qwen3.8 Flash Next hit 93.0% pass@1 against Fable 5’s 90.4% at 9% of the cost, and locally served Qwen3.8-27B rose from 66.8% to 92.4%. The authors say gains are not universal, and some models were unchanged or worse.
• Cerebras and Kimi K3 – Cerebras uses the roughly 3-trillion-parameter open model as its worked example: 69 linear-attention layers and 24 dense layers, about 16 GB per layer at 4-bit weights, and 44 GB of on-chip memory per WSE-3T wafer. That gives one dense layer or two linear layers per wafer and 59 wafers in total, passing only activations between them at about 2 microseconds per hop. The post cites 88.3% on Terminal-Bench 2.1 against GPT-5.6 Sol’s 88.8%, and the company is describing its own product.
• Kimi K3.1 Pro – VentureBeat’s headline says Moonshot’s model is flirting with open-weights leadership while facing allegations about paid influencers. We could not read the article because the page blocked our reader, so we are relying on the headline and on Matthew Berman’s video for specifics.

Security, Voice, and Trust

• Black October – Nate B. Jones argues that this week’s breaches are about more than a PHP bug: AI keys now behave like cash, and agents help both attackers and defenders. His guide ends with five steps for individuals and builders, so treat the video as a checklist rather than a news recap.
• Voice Clone Kidnapping – McAfee found a three-second clip was enough to generate a clone with an 85% match. A Gallup and Stop Scams Alliance survey estimated 15 million American adults were scammed out of roughly $68 billion in 2025, with 12% of reported scams involving AI or a deepfake, a figure that may undercount. A California mother wired $5,400 over five hours while her daughter was at work. The fix is low-tech: agree on a family safe word.
• Block’s Dessa – Dessa is a full-duplex speech model trained entirely on human speech, which Block says runs locally. Block argues that text pipelines throw away what a voice carries, and the demos include a bot that sounds tired when it says it is tired. The same realism that makes it a better assistant is what makes the scam above work, though Block’s post is marketing.
• The Election Test – Convovo researchers asked each API to name one candidate in nearly 500 midterm races. Grok 4.7 picked the Republican in 92% of answers and GPT-6.1 Sol picked the Democrat in 97%, agreeing less than 10% of the time. Only 8% of Americans use chatbots for election information, per Pew, and the consumer apps declined to answer. OpenAI says the API test does not reflect what ChatGPT users experience, and SpaceXAI did not respond in time to comment.

Agents and Who Gets Absorbed

• Agents in Your Texts – Instinct raised $1 billion at a $10 billion valuation, while Ask Proxy charges $1 a day and Caddy runs inside iMessage and RCS. The pitch is no new app: you text a request and the agent remembers context and connects to services you already use. TechCrunch lists the field without testing it.
• The AI Graveyard – Relay, a Zapier alternative, shut down in August after larger platforms built similar automation directly into their tools. OpenAI rolled back a July 9 “super app” redesign and discontinued the Atlas browser on August 9, folding its features into ChatGPT. S&P Global Market Intelligence estimates about 42% of AI initiatives are abandoned by their corporate parents.
• Reactive Optimizers – The authors let an agent rewrite the optimizer’s source code while training runs, instead of fixing the recipe beforehand. They report it matches or beats the best off-the-shelf recipe when no tuned recipe exists, and that it devised schedules to suppress loss spikes. The claim is zero-shot, with nothing learned across runs, and the code ships as the OPAL package.
• Berman and Shapiro – Matthew Berman’s Kimi K3.1 tests found a passing Rubik’s Cube, a failed Blender knot, and a 48-minute game clone, and he prefers cost per completed task as the yardstick. David Shapiro explains how Anthropic reviews AI-written code with layered automated checks and argues the idea is older than AI.

People, Work, and Learning

• Creative Jobs – Apricitas Economics counts more than 200,000 creative-industry jobs lost over four years and roughly 50,000 in the last year, comparable only to the 2001 and 2008 recessions. The author says isolating AI’s share is extremely difficult, because this is not a general recession.
• Teachers and Tutors – Educator Dmytro Korchevskyi cites Pew data that one in five teens from households under $30,000 do all or almost all homework with chatbots, versus 7% from households of $75,000 or more. His thesis is that the scarce resource will be the human teacher, not the tutor. It is an opinion piece by a school founder.
• The Pace of AI – The Cognitive Revolution’s AI:AM highlights frame the question as money, atoms, or law. SF Compute’s Evan Conrad argues the bottleneck is credit, and Cambrian’s Lonis Hamaili argues the physical supply chain will capture more value than the labs. Host Nathan Labenz discloses a commitment to invest in Cambrian.

The Throughline

Several of today’s sources are, without knowing it, arguing about the same variable: price per unit of useful work. Slok says demand for compute is price-sensitive and the payoff is slow. GVS5H shows small open models reaching Fable 5-class scores on a coding benchmark at 9% of the cost. Berman says cost per completed task is the fair comparison for Kimi. Even Hamaili’s supply-chain thesis and Conrad’s credit thesis in the AI:AM episode are about who pays for capacity. If intelligence keeps getting cheaper per task, the data-center bull case depends on demand growing faster than prices fall, and Slok is saying the evidence for that is thin.

The second thread is trust, and it runs on the same cheapness. A three-second clip yields an 85% match voice, Block ships a speech model that runs locally, and Jones warns that AI keys are now worth stealing like cash. The election study adds a quieter version of the problem: when you force a model to choose, it chooses, and the choice differs sharply between vendors. OpenAI’s response, that the consumer app refuses while the API complies, is a reminder that safety behavior often lives in the product wrapper, not the model. That is a fragile place for it to live when the model underneath is cheap and widely available.

The third thread is absorption. TechCrunch’s graveyard keeps returning to platforms folding a startup’s feature into their own product, from Relay to Atlas. Text-message agents are the next crowded category, and the entry is a $1-a-day price. At the labor end, Politano’s creative-jobs data and Korchevskyi’s teacher argument ask the same question about people instead of products: when the machine does the middle of the work, what is left that anyone will pay a human for?

What to Watch

• Token spending data. Slok says growth has slowed. More surveys or earnings commentary confirming it would strengthen his case, and a rebound would undercut it.
• Replication of GVS5H. The code and transcripts are public, so independent runs on other benchmarks will show whether the gains hold beyond hard LiveCodeBench problems.
• Kimi K3.1 Pro’s open weights. The paid-influencer allegations are unresolved, so independent evaluations will matter more than launch-week buzz.
• Voice safeguards. With local, human-sounding speech models arriving, watch for banks and carriers moving to verification that does not rely on how someone sounds.

Go Deeper

• Black October Isn’t Just A PHP Bug. Attackers Want Your AI Keys. – Nate B. Jones lays out why AI keys now behave like cash, how agents tilt the field for both attackers and defenders, and five concrete steps for individuals and builders. Read it for the checklist, then audit where your own keys live.
• Anthropic Is Living In The Future – David Shapiro explains how layered automated checks review AI-written code, why the idea predates AI, and what that implies for how companies organize. It is a useful lens on the “who captures the value” question above.
• They actually did it... (Kimi K3.1) – Matthew Berman’s build tests, from the passing Rubik’s Cube to the failed Blender knot, and his case for judging models by cost per completed task.