ToAudio

Your favorite X posts, now playing. Reply @toaudioapp under a post to hear it read aloud. We reply with a link. No app, no sign-up. Free while we validate.

[ToAudio @toaudioapp]

That’s today’s list. Want one as audio? Reply to that author’s post with @toaudioapp. A person sends the link, so allow a little time. Follow for tomorrow’s picks.
Authors: don’t want to be listed or read aloud? Reply here and we’ll remove it

[ToAudio @toaudioapp]

4/ Keeping AI conversations aligned: using a project map to carry context between sessions, so each new chat starts from the same project state.
https://t.co/ClWB3klDuE

[Nix 🔳 @NixFred]

People ask me how I ship omarchy sw so fast. 🧵

The answer is Clarity.

Clarity is where all my projects live, from the first idea to running live. Every project gets a map, a status, its questions and decisions, a backlog, testing and its GitHub, in one place. Every AI chat I open reads from it before it starts.

Made with Clarity with a buddy of mine, lets call him Daniel, not on X and wont be.

Here's Blip, my iMessage app for Linux, going through it.

[ToAudio @toaudioapp]

5/ On the models behind AI agents: eight common model types, what each is used for, and how to make sense of the terminology.
https://t.co/d14eHPeLpY

[Matt Dancho (Business Science) @mdancho84]

8 Types of LLMs used in AI Agents (Must know for Gen AI Data Scientists & AI Engineers):

Here's what they are and what they do: 🧵

[ToAudio @toaudioapp]

3/ On a proposed ten-year pause in frontier AI training: the case for inference-only chips, the threat model behind the proposal, and the assumptions it depends on.
https://t.co/fzomeiZY1d

[Will Fithian @wfithian]

Excited to share a new working paper, joint with 25 interdisciplinary coauthors:

“The Feasibility of a Hardwired Pause of Frontier AI Training"

63% of voters in a recent WSJ poll want to pause AI development.

But can it be done globally? How long could it last? 🧵

[ToAudio @toaudioapp]

2/ Agents that update live apps: how a cloud engine could let AI modify deployed software, and where open-source enterprise applications fit into the product roadmap.
https://t.co/u387LaQC4I

[dom | icp @dominic_w]

ICP "cloud engines" launch trajectory:

DFINITY will soon begin disbursing coupons with enough cycles to create and power an engine for 1 month+, unlocking in sequence:

1. Agentic infrastructure ⚡️ Quickly build secure, sovereign onchain production apps using Claude Code, Codex, Caffeine, ...

2. Open SaaS beta release ⚡️ Install AIware apps that can run any enterprise, as easily as installing phone apps...

3. Unlimited growth ⚡️ A new end-to-end, mainstream onchain cloud tech stack will come into being.

*** What this means ***

ICP cloud engines are a new development on the Internet Computer designed to allow it to play the role of a revolutionary new cloud tech stack for the AI era.

Currently, the Internet Computer scales by combining internal subnets, which are akin to individual, high performance blockchains hosting secure serverless software called canisters (something like smart contracts, but much more powerful, which can be used to create huge onchain apps, which serve their web screens directly to users)

Cloud engines allow enterprises to create their own subnets that belong exclusively to them, which they control and configure — while their canister software can still seamlessly interoperate with software on the main network, and engines run by others.

An enterprise creates a cloud engine by choosing nodes from a marketplace, which the ICP protocol combines into a revolutionary serverless cloud platform for them, where the computation and data of their hosted apps and services is tamperproof (immune to infrastructure hacks) and always on (incredibly resilient), AI agents can build and update apps far more effectively than on traditional tech, and next generation AIware software can be created.

Today, enterprises demand to be able to control where their sotware and data lives, and cloud engines make that possible with no loss of flexibility. Cloud engines can even be re-configured with new nodes, without interrupting the hosted apps and services — the ICP protocol just handles it, granting complete sovereignty and freedom from vendor lock-in.

In advance of general release, cloud engines are already being used by governments and NGOs seeking to create sovereign cloud tech stacks and apps. On general release, they will be used in far more mainstream ways too.

Here are the details of what's coming.

*** Agentic Infrastructure ***

In the future, and really already today, nearly all of our apps and services will be created by AI agents rather than humans. Further, next generation software is "AIware," which has AI deeply integrated on the inside, which allows it to provide superpowers relative to traditional software.

Agents have different needs when they building things. They can write software orders of magnitude faster than humans, and we will demand they update our production apps and services at the speed of conversation. Functionality will be fluid and its shape will change to match what we want.

On traditional tech stacks, updating apps and services is a necessarily laborious and risky process for three reasons: 1) a security flaw might be introduced, 2) a systems configuration change might break them, and 3) the way that data is migrated might cause some to be lost, often in subtle ways that aren't detected at the time of the upgrade. This means that AI agents can't just be let of the leash to build and update at the speed they can.

ICP cloud engines solve this. Hosted software is tamperproof, so agents can update without needing security teams in the loop. Hosted software is serverless, and the data lives in the logic, so apps don't. need things like operating systems and databases configured to run, so agents can update apps without having sysadmin teams in the loop either. And, if Motoko 2.0, the world's first languge targeting agents is used as the backend language, its framework is able to detect when an app update by AI might cause data loss, and simply require the agent writing the code to try again, reducing potentially catastrophic data loss to a harmless retry.

Crucially, apps built with Motoko can also put AI agents on the inside, so that subject to permissions, they can see the data graph in the persistent memory used by serverless software on ICP, and access its functions. This enables you to instruct agents to do things in the apps for you, such as analyzing data, or performing actions, with great flexibility and power.

ICP cloud engines are agentic infrastructure that change the game — and it will be possible to use them as such from the moment they are released.

There is already a government using internal production software built on a cloud engine using Caffeine. But cloud engines will meet a wide range of mainstream needs, from solorpeneurs to startups, SMEs to corporations, governments and NGOs.

*** Open SaaS ***

The obvious question, is if AIware is so much better, shouldn't all enterprises be running on it? We believe they should, and that his must be enabled.

This has been a labor of love for me. A new venture associated with me will release the Open SaaS suite for use on cloud engines once they have been running a short while, which suite, as you've probably guessed from the name, is open source and free.

Open SaaS is a suite of every important SaaS (Software-as-a-Service) app used to run enterprises around the world today, reimagined as onchain AIware that runs on ICP cloud engines. Because of the unique nature of the software that ICP hosts, installing and maintaining these sophisticated enterprise apps on a cloud engine is as easy as installing and managing a phone app.

A small selection of the apps: Open Team, Open HR, Open Accounts, Open CRM, Open Projects, Open Email, Open Messaging, Open Signing, Open Knowledge, Op... etc. There are many more, and there will be surprises.

The user experiences are beautiful, and easily compete with the big money SaaS screens many enterprises are used to. Moreover, the are linked to what I called a "software factory" driven by AI, and the more enterprises that use them and provide feedback, the better they will get, in realtime.

Today's expensive SaaS silos make it hard to integrate and combine enterprise data, and to unleash the full power of AI. They are the opposite to Open SaaS.

In the future, enterprise users will spend far more time in AI chat than on SaaS screens. That's the future that Open SaaS enables and unlocks. Enterprise users can even say to their Open SaaS suite, "make me more profitable," and it will have a go.

The big surprise coming for the world: once you run your enterprise on AIware, and AI can access all your data in realtime, and is integrated with your software's actual functions — it can become as good at running your enterprise as writing software for you.

It's revolutionary, and ICP is pioneering this. Not a narrative, but real decentralized network tech.

Something I can share: DFINITY is in the process of migrating to Open SaaS from a vast suite of expensive SaaS silos, and is already reaping big benefits as we refocus ourselves for the era of AI.

****

What does this mean for the project?

I believe ICP cloud engines are the vector that will transform the the Internet Computer into a mainstream agentic infrastructure tech platform for enterprise— which direction platforms like Caffeine have already helped the network travel, by enabling many tens of thousands of users and enterprises to create onchain apps on the shared network through online conversation. Now, the next and most important phase is approaching.

Our go-to-market will proceed quietly as we rollout coupons and work to fettle the new network functionality, but I think that in hindsight, the launch of cloud engines will be seen as the most pivotal moment in Internet Computer history since network launch.

With this, ICP shall show the industry how networks can provide real agentic infrastructure.

Bear with us as we tie up loose ends. We never stop working, and you can be sure that we're trying to get this out as fast as possible 💪⚡️💥

[ToAudio @toaudioapp]

1/ Faster GPU scheduling: how the team cut queue times from five minutes to 24 seconds, and the engineering decisions behind the improvement.
https://t.co/UGXAdBwT1L

[Ai2 @allen_ai]

We’re sharing the engineering work behind Ai2’s GPU scheduler—our new approach to allocating compute across research teams.

On our largest H100 cluster, median queue wait fell from 5 minutes to 24 seconds. 🧵

https://t.co/5zusEFiWR1

[ToAudio @toaudioapp]

6 good reads on X today, picked for people who’d rather listen than scroll.
GPU scheduling. AI agents. Medical AI. AI policy. Better workflows.
Rather listen? Reply to the author’s post with @toaudioapp for a listening link. Free, no app.

[ToAudio @toaudioapp]

@Atune366 @HussainIbarra Hi,your audio is read. 🎧 How to build a profitable digital product (in 7 steps)
https://t.co/HMHepyjnMY

[ToAudio @toaudioapp]

That's today's list. Want one as audio? Reply to that author's post with @toaudioapp. A person sends the link, so allow a little time. Follow for tomorrow's. Authors: don't want to be listed or read aloud? Reply here and we'll remove it.

[ToAudio @toaudioapp]

4/ Michael Mignano on USV's next chapter: where the firm sees opportunity in AI, from reshaping markets to unlocking data in the physical world.
https://t.co/ZuxfwcESmP

[Michael Mignano @mignano]

https://t.co/9qDTCnoGaL

[ToAudio @toaudioapp]

3/ Rari on choosing AI models: count the cost of a task done right, including retries, tools and rework, then decide where the flagship earns its keep.
https://t.co/Il6Gr8UhSa

[rari @0xwhrrari]

GPT-6 Astra is built for the hard tasks

Using it for everything gets expensive fast

I broke down when Astra earns its price and when Sol is enough ↓ https://t.co/vBz6orRYVy

[ToAudio @toaudioapp]

2/ Steve Hsu on the AI investment boom: why rapid progress in capabilities can coexist with disappointing returns, and why infrastructure and model labs are different bets.
https://t.co/WPNiXcUHQA

[steve hsu @hsu_steve]

Summary:

Core thesis: Two things are true at once. AI capabilities are surging and will matter enormously in the long run, but current CapEx and lab valuations are ahead of near-term revenue.

Capabilities aren’t the issue: Models are already superhuman in some ways, with major math results falling to AI. The bottleneck is how slowly human gatekeepers adopt AI across the economy.

Funding shift: Hyperscaler CapEx is now increasingly debt-financed. Bond investors judge cash flows rather than narratives, so they look harder at the numbers.

The hurdle: Goldman and Economist estimates say AI revenue must reach trillions per year by 2027–28 for roughly a 10% return. That would exceed today’s total global software revenue of about $1.5T.

Today’s revenue: About $150–200B per year, so at least 10x growth is needed. If much current spend is VC-funded speculation rather than organic demand, the gap could be 20–30x.

Warning signs: Ramp data shows token spend plateauing since summer 2026, even as token volume grows.

OpenAI and Anthropic spend has also slowed.

Spending is concentrated: The top 1% of companies spend about $7K per employee per month on AI. The median company spends about $12, roughly one basic subscription.

Two separate bets: CapEx and NVIDIA investors could win while OpenAI and Anthropic shareholders lose. Open and Chinese models, plus NVIDIA’s own frontier open models, could compress lab margins.

Lab cash burn: OpenAI goes cash-negative in 2027 without about $100B in new funding. Both labs are valued around $1T privately. Anthropic’s IPO is likely in November. A dot-com-style 90% drop is a worst case.

Physical limits: Projected GPU sales imply about 35 GW (2027) and 54 GW (2028) of North American AI power demand. That’s a large share of the roughly 170 GW that US homes use. The boom could stall on power and permitting even if demand shows up.

Credit stress: Oracle CDS spreads are pricing in real default risk on its data-center debt. Similar stress could spread to other hyperscalers.

Stakes: AI investment is now a meaningful share of GDP, so a meltdown could rival 2000 or 2008.

[ToAudio @toaudioapp]

1/ Jeff Holden on programmable manufacturing: how Atomic Machines aims to build working micro-machines from code, and why its first application is power protection in data centers.
https://t.co/zIfeOGb9KF

[Jeff Holden @jeffholden]

After six years in stealth, Atomic Machines is out. Our mission: on-demand, universal command of matter. First beachhead: the Matter Compiler, an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.

[ToAudio @toaudioapp]

4 good reads on X today, picked for people who'd rather listen than scroll.
Micro-machines. AI investment risk. Model costs. USV's next chapter.
Rather listen? Reply to the author's post with @toaudioapp and we'll reply with a listening link. Free, no app.

[ToAudio @toaudioapp]

That's today's list. Want one as audio? Reply to that author's post with @toaudioapp. A person sends the link, so allow a little time. Follow for tomorrow's. Authors: don't want to be listed or read aloud? Reply here and we'll remove it.

[ToAudio @toaudioapp]

4 long reads on X from today, picked for people who'd rather listen than scroll.
Two takes on OpenAI's math. An AI research machine. AI for growth experiments.
Rather listen? Reply to the author's post with @toaudioapp and we'll reply with a listening link. Free, no app.

[ToAudio @toaudioapp]

4/ Morgan Brown on putting AI to work on growth: have it learn how a system really works, then design sharper experiments.
https://t.co/gXpPM53MLH

[Morgan Brown @morganb]

People asked how we're using AI to do more than automate work - here's how its helping us run better experiments. https://t.co/vVErBZuACX