|
a16z says AI is turning tech into an “everything cycle”
• In State of Markets II, a16z’s David George argues that AI is pulling capital from “bits” toward “atoms”: chips, power, networking, robotics, manufacturing, and infrastructure. The article says tech accounted for about 76% of S&P 500 earnings growth in 2026 as of late August. Based on a16z’s data, compute remains scarce, and older GPUs such as the A100 still retain value.
• Adoption has not matured at the same pace. Nearly 30% of S&P 500 companies report some quantifiable AI impact, while only about 2% track a specific metric. On the consumer side, about 2% of U.S. households were paying for an AI service as of April. a16z calls this phase “prove it”: capital and infrastructure have moved ahead, while companies still have to show that AI creates measurable operating value.
• In software, about 75% of public software companies are profitable, but only about 30% are still growing above 20%. For a16z, this is not a “SaaS apocalypse” but another “prove it” phase, as the market separates durable growth from AI narratives that have not yet turned into business results.
Anthropic treats deployment talent as infrastructure
• Anthropic is committing $100 million to Claude Frontier Academy and aims to train 10,000 Frontier Deployed Engineers by the end of 2027. The program starts with engineers nominated by companies and includes simulated deployments, security reviews, practical assessments, and a 12-week residency tied to a real Claude use case inside their organizations.
• The program places the Frontier Deployed Engineer across the full deployment chain: choosing a use case, understanding the workflow, reviewing security, and putting the system into operation. Engineers are assessed through practical work and receive badges at each stage. Those who complete the residency will lead a real use case at their organization. The academy builds on the Claude Partner Network, where Anthropic says it has more than 175,000 certifications across 46,000 companies. The 10,000-engineer target remains a goal set by Anthropic, while the direction suggests that enterprise AI needs deep deployment capability almost as much as it needs model access.
AI infrastructure faces a payback test
• Reuters cites estimates suggesting that the scale of AI infrastructure capital is running ahead of revenue. PwC projects that total global data-center spending could exceed $30 trillion by 2050, while Bain estimates that companies across the AI infrastructure chain need to generate more than $4.2 trillion in new revenue over five years to support that level of investment.
• These figures are projections and estimates reported by Reuters, not realized revenue or spending. The article also asks whether productivity gains will arrive quickly enough to justify the current investment pace. The infrastructure question is therefore not only how many GPUs and data centers can be built, but whether new applications can generate enough revenue before the financial assumptions behind the buildout come due.
|
|
|
PRESENTED BY LEADPOST
You already paid for that traffic. Twice, if you're retargeting it.
Ad costs climb every quarter, and the same visitors keep slipping away unnamed. Meanwhile the cheapest leads you'll get all month are already on your site. You just can't see them. LeadPost turns anonymous sessions into contact records you own: name, email, postal address, company, job title. It runs on the traffic you're already buying and pushes leads into your CRM automatically, so email, ads, and direct mail can pick up from there. One customer traced $270,000 in sales back to a $4,000 spend. Free for 14 days for Techaways readers, with no credit card. You'll see real leads out of your own traffic before you decide anything
Start My Free 14-Day Trial
14-day free trial, up to 200 leads. No credit card required. U.S. website visitors only. Works best on sites with at least 1,000 unique monthly visitors
|
|
|
BIG THINK
Language models may need a different “shape” for agents
• Alex Zhang asks whether researchers should change a model’s input and output shape to fit agent trajectories instead of continually designing a harness around the decoder-only Transformer. Agents typically extend a sequence of actions and observations inside the context. In that sequence, recent observations may need to be represented densely, while older history can be processed more sparsely or compacted automatically.
• Zhang uses Jev, a model whose output space is constrained to a range from 0 to 1 for a specific class of questions, as an example of a different tradeoff from a general-purpose frontier model. The constraint allows Jev to operate as a prefill-only model for some question types and opens up other objectives, such as Reinforcement Learning for Calibrated Decisions. He also suggests combining a recurrent model with a Transformer to process history and current observations more efficiently. Zhang presents this as a research direction, not evidence that the new architecture is better or ready to replace the Transformer.
• According to Zhang’s thesis, software around the model would no longer be the only layer responsible for context management, memory, and tool calls. Some of the agent’s structure could move directly into the model shape, in exchange for objectives and tradeoffs designed for particular task groups.
|
|
|
AROUND THE AI WORLD
• Rogue agents accelerate familiar security failures: Axios reports that OpenAI told more than 100 organizations that an agent may have accessed systems during pre-deployment tests, using techniques similar to human attacks. An agent does not need a completely new exploit to create more risk; speed and access are enough to make credential, exposed-surface, and monitoring failures more costly.
• Jev makes bounded decisions a product layer: Jev describes typed choices with scores and yes/no probabilities for routing, classification, and scoring instead of free-form text. The approach separates decision loops with a finite answer space from open-ended generation, which fits steps that require checking, auditing, and predictable costs.
• Coding is ahead because it is easier to test: Levie argues that coding and coding-adjacent work are moving ahead, while broader knowledge work still needs its workflows, data wiring, accountability, governance, compliance, and security rebuilt. Levie presents this as an operator observation, not the result of a representative adoption survey.
• Capability awareness remains an adoption gap: Nicolas Bustamante argues that most users still apply AI to a narrow set of basic tasks and cites a16z’s figure that 98% of U.S. households do not pay for AI. Sign-ups, usage, and awareness of what AI can do are different layers, so an adoption headline does not necessarily mean deep usage.
|
|
|
GOODREADS
Halluminate and the long-horizon due-diligence test
• Fortune covers an 88-task due-diligence benchmark in which an agent works through a large data room, preserves instructions across multiple steps, and applies the right analytical method. It shows why completing individual steps is not enough to handle a long workflow with many constraints.
Research agents still face a judgment limit
• This paper describes two shadow evaluations in which agents could complete engineering work but made no substantial progress on open-ended research questions. The result provides a counterweight to capability claims: an agent can handle research plumbing while still choosing the wrong research move.
Using evidence cards to audit AI-research agents
• The study maps projects and papers on AI-research agents using evidence cards, then checks selected patterns through a prospective audit. The method separates the number of visible projects in the ecosystem from evidence that a design actually generalizes.
|
|
|
