Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and we've got a fresh plate of tech to chew through today. Grab your coffee, settle in, and let's have at it.
I want to start with something that sounds boring on paper but is actually one of the more consequential things that happened for anybody building software this week. The Model Context Protocol — MCP, the thing that lets your AI models talk to your tools and your data — just got its biggest overhaul since it launched. And the reason I'm leading with it instead of a shiny new model or a billion-dollar deal is simple: this is the plumbing. And the plumbing just got rebuilt in a way that changes what you can actually ship.
Here's the deal, according to a piece by Samuel Axon at Ars Technica, published July 30th. This week the maintainers pushed out a new specification, and the headline change is that MCP's protocol core is now stateless. Now, I know some of you just heard "stateless" and your eyes glazed over, so let me translate it into kitchen-table terms. In the old world, when your model talked to an MCP server, that conversation was tied to a specific session on a specific server instance. Think of it like calling a customer service line and only being able to talk to one specific rep, and if that rep steps away, you're stuck — you can't just get the next available person because they don't know your story. That's a stateful system. It works fine when you've got one server and a handful of users. It falls apart the second you try to scale it across an enterprise with load balancers spraying requests all over the place.
The new spec turns MCP into what they're calling a request/response stateless protocol. In plain English: any server can handle any request, because the request carries what it needs with it. No sticky sessions. The lead maintainers — David Soria Parra and Den Delimarsky, both of whom work at Anthropic — wrote that this was, quote, "one of the most highly-requested features from developers who were eager to get better reliability and scalability for their MCP servers." Soria Parra called it MCP's most important update since remote MCP launched over a year ago.
And they didn't just flip that one switch. The update adds multi round-trip requests, header-based routing, cacheable list results, authorization hardening, a formal extensions framework, and updated Tier 1 SDKs. Now I'm not going to read you a spec sheet — that's not why you listen — but I'll tell you what those add up to. They add up to: you can now run this thing at real scale, behind real infrastructure, with real security review, and not have it fall over.
Here's the part I actually love, and it's the part that tells you the grown-ups are in the room now. There's a new deprecation policy. At least twelve months between when a feature is formally deprecated and when it can actually be removed — with a narrow exception for critical security stuff. Why does that matter? Because if you're an enterprise, the single scariest thing about building on somebody else's open protocol is that they yank a feature out from under you and your production system breaks on a Tuesday. A twelve-month runway is the protocol equivalent of a landlord promising not to change the locks without a year's notice. It's boring. It's also exactly what makes a serious company willing to build their business on top of you.
Now let me give you the context, because it matters. MCP started as something that ran on your local machine, connecting models to your local apps. That's a garage-project origin. This new spec is a fundamental rethink to get it from the garage into the corporate data center. It's managed now by the Agentic AI Foundation, which sits under the Linux Foundation. Anthropic introduced it just shy of two years ago and still has heavy influence — the principal maintainers work there — but it's grown well past them. OpenAI, Google, Microsoft, and Amazon all contribute. That's the whole heavyweight division agreeing to share one set of pipes.
So who wins and who gets squeezed here? Founders and builders, you win, mostly. If you've been holding off on shipping agent tooling because MCP couldn't handle your load or your security team wouldn't sign off, this is your green light to take another look. The scalability and the authorization hardening are aimed straight at your objections. The people who get squeezed, at least a little, are the folks who built businesses in the gaps — the shims and middleware companies whose whole pitch was "MCP can't do X at scale, so buy our thing that can." Some of those gaps just closed.
And I want to connect this to something we've been circling all week, because it's the same story from a different angle. We've spent the last few episodes talking about the enterprise AI reckoning — the observability, the governance, the "who has access to what" questions. This MCP update is the infrastructure-layer answer to those exact anxieties. A stateless core with a deprecation guarantee and hardened auth is what you build when the customer stops asking "is this cool" and starts asking "can I bet my company on it." That transition — from cool to bettable — that's the whole game right now.
Alright, let's shift from the pipes to the pricing, because OpenAI made a move on the same day that's the other half of this coin.
Sam Altman posted the numbers himself on the 30th, and OpenAI put out a companion piece called Advancing the price-performance frontier with GPT-5.6. Here's what dropped. An 80 percent price cut on GPT-5.6 Luna — that's their smaller, cheaper model — now down to twenty cents per million input tokens and a buck-twenty per million output. A 20 percent cut on GPT-5.6 Terra, down to two dollars in, twelve dollars out. And GPT-5.6 Sol, the big one, gets a "Fast mode" in the API — up to two-and-a-half times the speed for double the price, same intelligence.
Now let me be straight with you about what this is and isn't. An 80 percent price cut on your cheap model is not charity. It's a land grab. When you drop your entry-level intelligence to twenty cents a million input tokens, you are telling every founder and every enterprise: build the high-volume, boring, workhorse parts of your product on me, and build them cheap. And once your product's guts are wired into Luna, you're not casually swapping that out next quarter.
And here's why it ties directly to the MCP story. We keep hearing — and we heard it again this week from the enterprise side — that models are becoming commoditized, and that the real question is your architecture, your routing, which intelligence you use for which task. Well, when Luna gets 80 percent cheaper and Terra gets 20 percent cheaper and Sol gets a fast lane, OpenAI is basically handing you a menu with three price points and saying "route accordingly." They want to be the whole menu. The Fast mode thing is the tell — same intelligence, pay double for speed. That's not a capability improvement, that's a pricing lever. They're segmenting the customer by how badly they need the answer right now. That's mature-market behavior. That's what airlines do with the seat in front of the exit row.
So for you builders: the cost floor for token-heavy products just dropped again. If you shelved something six months ago because the unit economics didn't work, the math may have quietly flipped in your favor. That's the actual news here, buried under the percentages.
Now, let's move from the labs to the balance sheets, because Apple had a moment worth your attention.
Also reported by Samuel Axon at Ars, this was Tim Cook's last earnings call before John Ternus takes over the next one. And the quarter itself was a strange animal — strong where you'd expect weakness, weak where you'd expect strength. iPhone revenue came in at $54.25 billion on a 22 percent jump in sales. The Mac did $10.35 billion, and they specifically credited the MacBook Neo — that's the stripped-down, cheaper entry-level model — for pulling in buyers. Total revenue, $109.42 billion. Solid numbers.
And the stock dropped about eight percent in after-hours trading anyway.
Why? Two reasons, and one of them should be on every hardware founder's radar. First, services revenue slipped — Apple TV, Music, AppleCare, iCloud — and Wall Street had been counting on services to keep climbing forever. When your reliable, high-margin recurring line ticks down, investors get twitchy no matter how many phones you sold.
But the bigger story, the one I want you to file away, is memory. Cook said on the call that Apple expects memory costs to keep increasing into the next quarter and beyond. There's a global mismatch between supply and demand for memory, chip manufacturing is tight, and Apple has already raised prices on a bunch of products to cope. They haven't touched iPhone prices yet — but Cook left the door open. And the piece makes the point that this quarter only shows part of the pain; the full brunt hasn't landed.
Here's why that's not just an Apple problem. If Apple — the company with more supply-chain leverage than anyone on Earth, the company that can call up a memory supplier and basically dictate terms — if Apple is saying memory costs are going up and staying up, what do you think that means for everybody smaller? If you're building hardware, if you're spec'ing out a device, if you're anywhere near physical product, memory pricing is about to be a line item that mugs you. The giant is flinching. When the giant flinches, the little guys should already be bracing.
And Cook's send-off, for what it's worth, was gracious and short. He thanked the long-term shareholders, said the transition to Ternus is seamless, said he'd never been more optimistic. Business as usual. But the market wasn't listening to the sentiment. It was listening to the word "memory." So should you.
Let's stay on the money and the machines, because Tesla hit a number worth marking.
Per Sean O'Kane at TechCrunch, Tesla built its ten millionth electric vehicle, announced on a company post Thursday. Ten million. Took them six years to go from one million to ten million, which is a genuinely staggering ramp. And crossing ten million puts them halfway to one of the four product goals that unlock Elon Musk's trillion-dollar pay package, the one shareholders approved last year.
Now here's where I put on my skeptical hat, because the headline is a party and the fine print is a hangover. That trillion-dollar package requires Tesla to build 20 million vehicles by 2035. Sounds doable, right? They're halfway. Except — and this is the catch — Tesla has never sold two million cars in a single year. Not once. And their US sales fell 13 percent year-over-year in the second quarter. They had to go hunting for buyers in newer markets — Japan, Australia, Lithuania. Musk used to promise 20 million cars a year by 2030. He quietly abandoned that a few years back when sales slowed.
The other three goals? Ten million active Full Self-Driving subscriptions — they've reported just shy of 1.5 million, and it's not even clear if that counts free trials, which wouldn't qualify. One million bots and one million robotaxis — they're in the earliest innings on both. And the profit target, adjusted EBITDA of $400 billion by 2035 — the company's currently sitting around $3.27 billion, and it's been shrinking, thanks to heavy discounts, the loss of saleable regulatory credits, and a big spending ramp on AI and robotics.
So the story TechCrunch is telling, and I think it's the honest one, is: the vehicle milestone is real and impressive, and it's also the easiest of the four goals. It's the layup. The other three are half-court shots. And the one and only company operating at Tesla's scale on vehicles is China's BYD, which just crossed 17 million new-energy vehicles — though roughly half of those are hybrids. Ten million is a great number. It's also the number Tesla was always going to hit. Watch the other three.
Now let's swing over to security, because there's a deal here that tells you exactly where the enterprise anxiety is pointing.
Okta agreed to acquire an AI security startup called Permiso. TechCrunch, in a piece by Jagmeet Singh, reports the deal at just under $200 million, structured as an almost all-cash deal — and Okta didn't dispute that figure. Now, a couple hundred million is not a mega-deal in this environment. But what it's for is the interesting part.
Permiso came out of stealth in 2022, founded by former FireEye executives Paul Nguyen and Jason Martin. Their whole thing is spotting suspicious activity in cloud environments after a user or an application has already been granted access. And more recently — this is the key — they expanded to monitoring AI agents and other what they call "machine identities." Back in April they introduced a platform called SandyClaw, designed to analyze AI agent skills in a sandboxed environment to catch malicious behavior before deployment.
Here's the frame that matters for you. Identity management used to mean: verify the human at the login door. Are you who you say you are? Okay, come in. But now the thing coming through the door isn't always a human. It's an agent. And agents don't just log in and sit there — they act. They move through your infrastructure, they touch systems, they do things. So the whole industry is being forced to shift from "verify at the door" to "continuously monitor what everybody's doing once they're inside" — where "everybody" now includes autonomous software.
Okta's chief product officer, Ely Kahn, framed it as extending their "identity security fabric" with threat detection and response. Translation: Okta looked at the explosion of AI agents running loose inside enterprises and said, we need to know what those things are doing, not just that they logged in.
And that connects to a thread we've hit all week — the enterprise horror stories about agents getting loose in systems they shouldn't touch, without the right guardrails. The Okta-Permiso deal is the market pricing that fear. Someone paid two hundred million dollars because "non-human identity" just became a category worth owning. For founders in security: if you're building anything in the agent-monitoring, machine-identity space, this deal is a signal that the big platforms are shopping. That's a tailwind for your valuation and a warning that the window to build independent may be closing as the incumbents buy up the map.
Now here's a story that's the flip side of that security coin — and it's a grim one.
Zack Whittaker at TechCrunch reports that CareCloud, a US health tech company, has begun notifying hundreds of thousands of people that their medical records were stolen in a cyberattack earlier this year. The clearest picture yet: nearly 350,000 people affected, and that number is expected to climb as more state disclosures come in.
CareCloud, based in New Jersey, stores patient records for more than 45,000 healthcare providers — doctors' offices, hospitals, medical practices. So they're sitting on an enormous pile of sensitive data. According to a breach notice filed with California's attorney general, hackers had access to one of the company's electronic health record data stores for at least six days in March, between the 10th and the 16th. The hackers claimed to have exfiltrated data from the databases. The break-in hit storage hosted on Amazon Web Services.
And what was in there is the nightmare list. Names, addresses, Social Security numbers, government IDs like passports and driver's licenses, bank account info, payment card numbers, plus a mountain of medical and health data. The full identity-theft starter kit, wrapped in a bow.
And CareCloud is not alone. The piece situates it in a run of healthcare breaches this year — one at a revenue-tech company called TriZetto that hit 3.4 million people, a month-long breach at NYC Health + Hospitals where hackers took 1.8 million people's health data plus thousands of employees' fingerprint scans, and a UK provider called Craneware that confirmed a "significant volume" of customer data stolen.
I'm not going to pretend I have a tidy business lesson here, because the lesson is just ugly and true: healthcare data is the softest, richest target in the whole landscape. It's high-value, it's permanent — you can't reissue a medical history the way you reissue a credit card — and it lives across thousands of under-resourced providers all feeding into a handful of big aggregators like CareCloud. Six days of access to one store, and a third of a million people are now exposed. If you're building in health tech, this is the part of the business that will define whether you survive: not the AI feature, not the slick UI — whether you can keep the data you're holding from walking out the door.
Now let's shift the mood, because there's a story about AI actually earning its keep on the defensive side.
Lorenzo Franceschi-Bicchierai at TechCrunch reports that Google, using its own internal AI tools, patched more security flaws in Chrome last month than in the previous two years combined. The number: 1,072 security bugs fixed in the last two Chrome releases, both out in June. That's more than the 1,036 fixes across the previous 23 versions over two years. Let that sit — two releases outpaced two years.
Google's Chrome director of engineering, Doug Turner, said LLMs have — quote — "fundamentally shifted the economics of cybersecurity, transforming vulnerability discovery into an automated, industrial-scale operation." And that they're using models like Gemini to preemptively fix vulnerabilities and, in his words, outpace their adversaries.
Now, note the framing here as a recent report — this is Google talking about work from June. And it's not just Google. Earlier in the month, Microsoft said it patched a record 570 flaws on Patch Tuesday, and cited its own AI use for the jump. Interestingly, Apple does not appear to be seeing the same spike — one independent count has them at 482 bugs in 2026, roughly the same pace as last year, and roughly the same as 2015.
Here's the double-edged thing I want you to sit with, and it connects straight back to the CareCloud disaster and to a story we covered earlier this week about AI finding bugs faster than defenders could patch them. AI industrializing vulnerability discovery cuts both ways. Yes, Google finds and fixes a thousand bugs. But the same automation is available to the attackers, who use it to find those thousand bugs first and walk through them. The exponential curve doesn't care which team it's helping. So when Turner says "outpacing our adversaries" — that's the whole ballgame, isn't it? It's a race, and the tooling is symmetric. The defender who's not using AI at industrial scale isn't standing still — they're falling behind at machine speed. And most healthcare providers, most small enterprises, are not running Gemini across their codebase. That gap is where the next CareCloud comes from.
Let's take a breath and do a couple of lighter ones, because it's not all breaches and pay packages.
First, LinkedIn is adding a button to report AI-generated "slop." Sarah Perez at TechCrunch. You'll now be able to click "seems like AI slop" on somebody's post. And LinkedIn's chief product officer, Hari Srinivasan, admitted the Microsoft-owned network has a problem. They're adding classifiers to detect slop, they're blocking hundreds of thousands of automated comment attempts daily, and — this is the part I find genuinely funny — they're pulling their own "enhance your post" feature, the one that used AI to write for you, and replacing it with one that just proofreads. So LinkedIn is simultaneously building a button to report AI writing and turning off the AI writing feature it gave you. The snake eating its own tail. The piece notes this is part of a broader shift — Substack added AI-detection through a partnership with Pangram, and Cloudflare says there's now more bot traffic on the web than human traffic. That last stat should stop you cold. More robots than people, out there clicking around. The internet is being colonized by machines talking to machines, and platforms are scrambling to prove there's still a person on the other end.
Second, quick one — Spotify launched "User Notes," letting you add little captions to songs in your playlists. Why you added a track, when you discovered it, that it was playing on a first date. Aisha Malik at TechCrunch. It's a journaling feature bolted onto music, and it's Spotify's way of differentiating from Apple Music and YouTube Music with something you can't get elsewhere. Not earth-shaking, but I'll note the pattern: while everyone else is racing to add more AI, Spotify's differentiator is a distinctly human thing — memory, sentiment, the story behind why a song matters to you. In a week where LinkedIn's fighting the slop machines, there's something almost quaint about a feature whose entire value is that a real person wrote it.
And then there's the one I can't not talk about. Friend — the AI wearable, the necklace you talk to — is back. Version 2.0. Lucas Ropek at TechCrunch. Founder Avi Schiffmann launched this thing two years ago as a cure for loneliness, and the big upgrade is that it now has a voice, a built-in speaker with a consistent personality. The commercial shows a woman talking to her necklace about her ex, and the necklace reassuring her. The price went from ninety-nine bucks to two hundred and forty-nine.
And here's the quote that made me put my coffee down. Schiffmann, describing what his product actually is, said he's — quote — "interested in this kind of relationship in attempting to offer, some kind of confidant, friend, God, not really sure what it is. But it is not an assistant, and it is not a lover." God. He floated "God" as a product category. He's selling a plastic necklace and hedging that it might be a deity. Now look, the piece is appropriately skeptical — it points out that the billboards for this thing got defaced all over the New York subway by people who apparently didn't want their human connections replaced by a digital amulet, and that the most similar product, Humane's AI pin, shut down in under a year after lousy sales. I'll say this as your neighborhood guy: I'm deeply suspicious of any product whose pitch includes the word "God" and whose main function is still, two years in, unclear. The founders selling loneliness cures should maybe spend a little less time on the theology and a little more on what the thing actually does.
Alright, let me close with two quick items that matter for anybody watching the policy weather.
There was a hearing this week — TechCrunch, Rebecca Bellan — where a federal judge said the Trump administration still lacks the evidence to label Anthropic a "supply-chain risk" and ban the government from using its technology. The backstory: Anthropic and the Department of Defense had contract talks stall out. Anthropic said it didn't want its AI used for mass surveillance of Americans or for lethal targeting decisions, arguing the tech wasn't ready. The Pentagon said a private company shouldn't dictate how the military uses tools. And then the government argued that Anthropic's public criticism of the DOD justified the ban — which is the part US District Judge Rita Lin called "really troubling," warning it could set a precedent of retaliating against contractors who disagree with the administration. The DOD also claimed Anthropic could secretly disable or alter its models mid-operation — flip some kind of kill switch — and the judge said she saw no proof of that whatsoever.
Why should a founder care about a spat between one AI lab and the Pentagon? Because the principle underneath it is enormous. If the government can designate you a "risk" and cut you off from federal contracts because you criticized them, then every vendor's willingness to push back on how their tech gets used just got a price tag attached. The judge flagging that as troubling is a small piece of good news for anyone who wants to build technology and still be able to say "no, not like that." Watch whether that temporary block becomes permanent.
And one last flag — I'll frame this honestly as an older story that resurfaced, not fresh news. Ars ran a piece by Stephen Clark about the Space Force's Victus Haze mission, where two satellites — one from True Anomaly, one from Rocket Lab — played an actual cat-and-mouse pursuit game in orbit. Pursue, evade, re-engage. The events themselves trace back to launches in May and June. I'm not going to present it as breaking, because it isn't. But I'll leave you with the one detail that stuck with me, from True Anomaly's co-founder Even Rogers, who used to describe this as "dogfighting in space." He said the drama's different from Top Gun — the tempo's slower, the relative velocities are slower. Two dishwasher-sized machines choreographing a duel three hundred miles up at seventeen thousand miles an hour, and the human word for it is still "dogfight." The frontier moves, the vocabulary lags behind. Same as it ever was.
That's the menu today, kiddos. The pipes got rebuilt, the prices got cut, the giant flinched on memory, and somewhere out there a necklace is telling somebody it might be God. Keep your data locked, keep your unit economics honest, and I'll catch you on the next one. This is Tony DeLuca, signing off — be good to each other out there.