Apple is reportedly paying Google around $1 billion per year, in a multi-year deal, for a customized 1.2-trillion-parameter Gemini model that will power Siri AI in the cloud. The new Siri is designed to be more conversational, context-aware, and deeply integrated into the operating system. It can understand what’s on your screen, interact with apps, answer questions about content on your device, assist with messages and calendar management, and maintain conversations across iPhone, iPad, and Mac. Users can interact with Siri through voice or text, and Apple says all processing occurs either on-device or through its privacy-focused Private Cloud Compute infrastructure.

Beyond Siri, Apple introduced AI enhancements throughout its software ecosystem. The Photos app gains new generative editing tools, including improved object removal, image expansion, and a feature called Spatial Reframing that lets users alter the apparent camera angle and composition of existing photos. AI-generated edits will be tagged with hidden SynthID watermarks. Safari now includes AI-powered tab organization, website monitoring for events such as price drops or ticket availability, and automatic password remediation for compromised credentials.

The new Siri is available to developers immediately and will enter public beta later this year. Initial availability will be limited to English, with support for additional languages planned. Regulatory constraints will delay launch on iPhone and iPad in the European Union and prevent launch in China altogether. Some of the most advanced on-device AI capabilities will require Apple’s latest hardware, including the iPhone Air, iPhone 17 Pro, M4-equipped iPads, and M3 Macs with at least 12GB of memory.

Apple positioned its slower approach to AI as deliberate rather than reactive. During the keynote, software chief Craig Federighi argued that many companies are pursuing AI for its own sake, while Apple is focused on creating AI that is useful, private, and centered on the user. Still, many of the newly announced capabilities closely resemble features already available through Google Gemini, ChatGPT, Claude, and other AI platforms.

The press is saying, “Apple finally caught up,” but we won’t know if that’s true until we start using the new Siri/Gemini hybrid. That said, I’m really excited about this. Maybe, possibly, there’s a chance that someday soon, Siri won’t suck.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Apple Outsources Siri’s Brain to Google originally appeared here on Shelly Palmer

What You Actually Buy

The three largest AI providers in the world publish their prices in tokens. A token is a model-specific unit of text representation. Different models tokenize identical content differently, which makes token counts a poor proxy for business value. Anthropic charges $5 per million input tokens for Claude Opus 4.8. OpenAI charges $5 per million input tokens for GPT-5.5. Google undercuts both at $2 per million for Gemini 3 Pro. On the published rate card, the two American flagships are priced in lockstep and Google looks like the bargain of the category.

Gemini looks less expensive to use. But tokens are an artifact of how transformer models bill internally, with no inherent relationship to the value the buyer extracts. Two identical tasks on two identical inputs can use wildly different token counts depending on the model’s reasoning style, prompt processing, tool calls, and verbosity defaults.

Anthropic says the new tokenizer in Opus 4.8 can use up to 35% more tokens for the same fixed text than earlier Opus models. That means the effective cost of the same job can rise by as much as a third without the published price changing by a penny. A Gemini task may run through twice as many tokens as a Claude task and arrive at the same total bill. The token price tells you almost nothing about the actual transaction. This would be a minor accounting problem if everyone agreed on a better unit.

The Candidate Units

The candidate units of intelligence currently in circulation are, roughly:

Tokens. What the API meter shows. Technically precise. Tokens are economically meaningful to the seller but only indirectly meaningful to the buyer for the reasons stated above. Favored by foundation model companies because tokens let sellers advertise plunging prices for capability they were not actually pricing in the first place.

Compute, measured in FLOPs or GPU-hours. What it costs to run the model. The unit Nvidia loves, because Nvidia sells the substrate. Useful for capacity planning, useless for procurement.

Task-completion length, measured by something like METR’s benchmark. What the model can actually do, expressed in how long it would take a competent human. METR’s data shows the longest task an AI can reliably complete is doubling every four months, down from every seven a year ago. Arguably the most economically meaningful unit currently available. It is also the hardest to standardize across vendors, which is why no vendor has rushed to adopt it.

Agent-hours or completed agentic tasks. What an autonomous AI system finishes on your behalf, measured in business outcomes rather than model calls. The unit Anthropic and OpenAI are starting to position around, because it lets them sell solutions rather than API calls. It is also the unit Salesforce, ServiceNow, and Microsoft want to own, because they already sell business outcomes by the seat and have decades of experience charging for them.

Intelligence-per-watt. What you get per kilowatt-hour of inference energy. The Nvidia and hyperscaler unit. This one matters more than people realize when the marginal cost of intelligence converges on the marginal cost of electricity, which is Sam Altman’s well-known prediction.

Each candidate unit advantages a different vendor. Each vendor knows this. The competition to define which unit becomes the industry standard is more strategically important than the competition to ship the next model release. Models obsolete in quarters. Units, once adopted, last decades.

Whose Unit Is It?

AI vendors sell tokens, agent runs, context windows, and benchmark scores. Enterprises buy contracts reviewed, tickets resolved, campaigns launched, code shipped, and revenue generated.

A CIO does not care how many tokens a task consumed. A CIO cares whether the task got done, how well it was done, and what it cost compared to the alternative.

That distinction is where the real unit war will be fought. Every vendor wants the market measured in the units that make its products look most valuable. Every buyer wants the market measured in the units that map directly to business outcomes.

The company that successfully defines the buyer’s unit will have far more pricing power than the company that merely defines the technical one.

The Unit Will Define the Market

In every previous infrastructure category, the company that defined the unit captured the market for a generation. Large-scale electricity markets accelerated after standardized metering and billing units emerged. Network bandwidth standardized on the megabit. Amazon turned compute into a broadly consumable utility product when it standardized on selling EC2 hours. The unit is the focus of the contract.

The AI unit war is already underway. Anthropic recently disclosed that a substantial majority of code merged into some internal repositories was AI-generated. Most people heard a capability story. I heard a unit story. The implicit benchmark is “lines of merged code per engineer per day.” If that becomes an industry-standard productivity unit, every other vendor will have to match the metric or argue against it, and both outcomes favor Anthropic.

OpenAI’s push around agents that can handle multi-hour tasks is a unit move. If “completed agentic task” becomes the unit, OpenAI’s investment in orchestration becomes the moat. Google’s integration of AI Mode into Search is a unit move. If “answer engine query” becomes the unit, Google’s distribution becomes the moat. Salesforce’s Agentforce pricing is a unit move. Microsoft’s Copilot-per-seat pricing is a unit move. Nvidia’s intelligence-per-watt benchmarking is a unit move.

Six different units, six different attempts to measure value, each uniquely beneficial to their respective creators. None of them are necessarily wrong. They are simply optimized for different buyers, sellers, and economic incentives.

Defining Your Unit

We’ve been working with our clients to define a Price Per Intelligence Unit and using that definition to build a way to evaluate costs on an apples-to-apples basis. Not in the abstract. Concretely:

• A marketing brief processed end-to-end with brand-aligned output?
• A customer service ticket resolved without human escalation?
• A legal contract reviewed for ten specific risk flags?
• A code review completed at a defined quality bar?
• A sales lead qualified against your eight-point rubric?

You can do this yourself. Pick the unit that maps to a job your business actually performs. Measure it in dollars per unit today, against the human-only baseline. Track the trajectory. The price per intelligence unit, on your metric, is falling fast, probably faster than your planning models assume. The compounding effect across a two-year capital cycle is enormous, and it will probably change the answer to several “build vs. buy” decisions on your roadmap.

When vendor evaluations finally happen, compare vendors on your unit. Make them justify their pricing inside your framework. The conversation changes when the buyer brings the unit definition. You stop being a tokens customer and start being a procurement counterparty the vendor has to negotiate with on your terms.

The Bigger Picture

The price per intelligence unit is collapsing across every reasonable definition of the unit. Eight times more code per day. Four-month task-length doubling. Four times cheaper per token year over year on roughly equivalent capability. Pick any unit and the trajectory is the same: down and to the right, on a curve few commodities have ever traced before. This is the deflation that ate the cost of computation in the 1990s, the cost of storage in the 2000s, and the cost of bandwidth in the 2010s, happening faster and across more dimensions simultaneously.

Model quality benchmarks are the measurements getting the big headlines, but a universally agreed-upon unit definition is going to be the key to quantifying the enterprise value of AI. Start by defining your own.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Price Per Intelligence Unit originally appeared here on Shelly Palmer

The most interesting data point comes from independent research organization METR. Their benchmarks show the length of tasks frontier AI models can complete reliably is now doubling every four months, down from every seven months just a year ago. In practical terms, AI systems that could handle tasks measured in minutes two years ago can now complete tasks that take hours.

In one example, Claude generated approximately 800 bug-fix pull requests that reduced a class of API errors by a factor of one thousand. The engineer who supervised the work estimated it would have taken a human several years to achieve the same result.

Back in December, OpenAI published its thoughts on recursive self-improvement (AI systems capable of helping build future AI systems). Anthropic is now describing something very similar: AI is increasingly helping develop AI, and they expect the pace of improvement to accelerate.

Applying this to your own business will require a few assumptions. Assume Anthropic’s results are 50% better than you can do without their engineering team and their tools. How will headcount and budgets have to change if your tech teams become 4x more productive? How will workflows have to change as coders transition from “writing the thing” to “reviewing the thing a model wrote”? And, of course, how will you create a culture of continuous adaptation so that your organization does not fall behind?

Most of my time is spent helping business leaders work through questions like these. Please feel free to reach out if you’d like to chat.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post When AI Builds Itself originally appeared here on Shelly Palmer

The change follows action by the UK’s Competition and Markets Authority (CMA), which required Google to provide publishers with AI-specific controls after designating the company as having Strategic Market Status. The CMA first proposed the measure in January, arguing that publishers should have a stronger position when negotiating how their content is used by AI systems.

The new control creates the appearance of choice, but the economics remain challenging. Publishers can opt in and allow their content to appear inside AI-generated answers that often satisfy a user’s query without generating a click, or they can opt out and disappear entirely from Google’s AI products while those products continue to occupy premium real estate at the top of the search results page.

Roger Lynch, CEO of Condé Nast, recently said he told his teams to “assume there’s no search.” He expects search traffic to become a single-digit percentage of total traffic. That’s a remarkable statement from one of the world’s largest publishers. If you depend on referral traffic, Google has now confirmed where search is headed. Plan accordingly.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Google Gives Publishers an AI Opt-Out originally appeared here on Shelly Palmer

A Corporate Claw is a safe-for-work version of OpenClaw (the Agent formerly known as Clawdbot and Moltbot), also known as a “persistent AI agent.” Claws connect to your email, calendar, files, and other data. They remember what you do and can continue working on your behalf even when you’re not around.

Microsoft also introduced Microsoft Execution Containers (MXC), a Windows-based sandbox designed to contain agent behavior. MXC limits what an agent can see, touch, and execute, reducing the risk that an autonomous system gains unrestricted access to your machine.

This is critically important because an agent with access to your inbox, files, and the authority to execute code creates a very large attack surface. OpenClaw-style agents are powerful, but they require serious security controls. Microsoft has joined NVIDIA and others who have acknowledged that while autonomous agents may be the future of productivity, containment and governance need to arrive on day one.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Scout: Microsoft’s New OpenClaw Collab originally appeared here on Shelly Palmer

Because the filing was submitted confidentially, there are no financial details yet. We don’t know the size of the offering, the proposed valuation, the float, the lock-up provisions, or the timeline for a public listing. Those details will become available only if and when Anthropic publicly files its registration statement.

What we do know: Anthropic is the first major frontier AI company to publicly announce that it has begun the IPO process. More importantly, Anthropic is a Public Benefit Corporation. Unlike a traditional corporation, its charter requires directors to balance shareholder interests with the company’s stated public-benefit mission. In Anthropic’s case, that mission includes the safe and responsible development of advanced AI systems.

If Anthropic goes public, there will be some obvious questions. For example: How will public shareholders value safety investments that may reduce short-term profits? How will analysts evaluate decisions driven by mission rather than margin? How much patience will the market have for a company whose governance structure was designed to resist some of the traditional pressures of public ownership?

Those questions have no precedent in the AI era.

The foundational model builders are beginning the transition from private capital to public markets. Once that happens, every enterprise that depends on these models becomes part of a much larger financial ecosystem. AI strategy, capital allocation, and market expectations will become increasingly intertwined.

We don’t yet know when Anthropic will begin trading, or what investors will ultimately decide the company is worth. We do know that the race to bring frontier AI companies to the public markets has officially begun.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Anthropic Just Started the AI IPO Race originally appeared here on Shelly Palmer

The Toolkit Everyone Has and Almost Never Uses

The tools for critical inquiry are well established. The devil’s advocate (Advocatus Diaboli, a role the Catholic Church formalized in 1587 to vet candidates for sainthood) argues against the proposal. The Socratic method exposes buried assumptions through disciplined questioning. Dialectic, from Plato through Hegel, seeks a higher-order answer by testing thesis against antithesis. Most “best practices” documents codify some version of all three.

In the real world, however, they are rarely applied as designed. An SVP playing devil’s advocate to the CEO is structurally compromised before they say a word. They know what their bonus depends on, who championed the proposal being discussed, and who will remember being challenged in front of the boss. An executive who asks an uncomfortable Socratic question quickly earns a reputation for being difficult.

Disagreement is socially expensive. That cost hollows out every well-intentioned adversarial-thinking technique. The next question, the one that would have changed the answer, almost never gets asked because nobody in the room can afford to ask it.

Why AI Sparring Partners Work

Current LLMs do not think. They pattern-match. They have no judgment, no taste, and no scar tissue. On a good day, they’ll produce a competent answer assembled from patterns in their training data. That’s a serious limitation.

But they do have one extraordinarily useful property: they don’t care. They have no career to protect, no boss to impress, and no relationships to manage at the next industry dinner. They don’t get tired, bored, defensive, or angry. They don’t read the body language that tells the rest of us when to stop. They’ll challenge an assumption, pursue a line of questioning, or explore an alternative path for as long as you ask them to. This is not intelligence. But it is useful.

Practical Applications

Two quick examples from the adversarial environments I’ve created for myself. One is an AI Board of Directors with five personas: a F500 CFO, a COO operator, a cynical activist investor, a technical risk officer, and a chair who synthesizes. The other is a Synthetic Focus Group built from our actual newsletter subscriber data.

After months of running both, the lesson is uncomfortable. My human advisors give better individual answers. The bots give better processes. They push past the polite stopping point. They challenge premises my advisors share with me because we all work in the same industry. They have no professional history that makes certain topics awkward to raise.

The Model Is Not the Answer

AI is not a replacement for human debate. Your human advisors carry institutional memory that is in no training corpus. They have lived through prior cycles of the same mistake and remember which executive pushed which doomed initiative. They can be held accountable in ways a bot cannot. They have long-running relationships that let them tell you the things they would never put in writing. None of that is reproducible in a system prompt.

The advice AI generates is often wrong, shallow, and suspiciously overconfident. Treating raw AI output as definitive is malpractice.

That said, a well-built sparring partner forces you to think harder about your own reasoning. It exposes the move you skipped, the assumption you smuggled in, and the question you flinched away from. The value lives in what changes in your head before you write the memo, not in the words the model gives you back.

How to Build Your Own

You do not need an engineering team. You can build a credible synthetic sparring partner with one prompt:

“You are a panel of four advisors reviewing a decision I am about to make. Each of you has a defined perspective: a skeptical CFO who has seen this story before; an operator who has to execute whatever we decide; a customer who will be on the receiving end; and a regulator looking for the failure mode. None of you has any relationship with me. Your job is to build the strongest case AGAINST the decision I am about to describe. Do not balance. Do not hedge. Argue as if your bonus depended on winning. Then ask me the three questions I have not yet addressed.”

Run the prompt before you write the memo, the order is the discipline. If you state your conclusion first and then ask the model to stress-test it, the model will mostly help you confirm it. That’s laundering, not reasoning. It’s why so many well-intentioned prompts produce useless answers.

A synthetic focus group works the same way. Build five or six personas from whatever customer, subscriber, or constituency data you actually have. Ask them to react to the headline or campaign concept before you fall in love with it. Have them argue with each other.

A critical-thinking rubric works the same way too. Write down the five questions a hostile reviewer would ask (or ask the model to generate them) and force yourself to answer all five before you draft.

This Does Not Replace What You Already Know How to Do

I teach two structured-thinking techniques in my class at Newhouse. The Famous “P&G Memo” (Idea, Background, Recommendation, Rationale/Discussion, Next Steps) forces brevity and structural clarity. The Amazon “working backwards” press release forces you to write the customer-facing announcement before you build the thing, surfacing every assumption about who actually cares. Both have helped executives think more clearly for decades. Neither is going anywhere.

An AI sparring partner is not a substitute for structured thinking. It’s an amplifier. Write the P&G memo. Draft the Amazon press release. Do the work. Build the argument. Then ask the machine to attack your assumptions, expose the weak spots, and identify the risks you missed. The value of structured disagreement compounds quietly. A flawed strategy gets fixed before launch. A bad assumption gets exposed before it reaches the board. A costly mistake never happens.

Twenty-five hundred years after Socrates complained that nobody in Athens was willing to think hard enough, the tools to help us think harder are hiding in plain sight.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post The Next Question originally appeared here on Shelly Palmer

The company is pushing “honesty” as the new feature. Anthropic says Opus 4.8 is about 4x less likely than 4.7 to let flaws in code it wrote pass unremarked, and early testers say it flags its own uncertainties and makes fewer unsupported claims.

The sticker price is flat at $5 per million input tokens and $25 per million output. Fast mode runs at 2.5x speed and is now 3x cheaper than it was on previous models. Databricks reports its Genie product reasons over PDFs and diagrams at 61% lower token cost than on 4.7. More capability, lower cost, same price. That is deflation doing its work.

Alongside the model, Anthropic shipped effort control (you pick how hard Claude works on a response) and dynamic workflows in Claude Code, where one session plans a job, fans out hundreds of parallel subagents, verifies its own output, and runs migrations across hundreds of thousands of lines of code. The atomic unit of work moves from “ask a question” to “hand over a project.”

Anthropic shipped a modest, safe model while openly sitting on a more capable one it will not release. Project Glasswing members already use “Claude Mythos Preview” for cybersecurity, and Anthropic is holding the more powerful class back until its cyber safeguards are ready, which it expects in the coming weeks.

Frontier model innovation is accelerating, but the models the public can buy today are deliberately a step (or two) behind the models that exist.This raises the question that keeps me up at night: if this is what’s going on with LLM-based reasoning engines, what will happen when a lab achieves AGI?

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Claude 4.8 is Here originally appeared here on Shelly Palmer

This is the same playbook Amazon used to build AWS two decades ago. Build technology to solve your own problems, package it as a service, then sell it to everyone else.

Retailers running ASA will run their AI shopping on Amazon’s rails (Bedrock, AgentCore, OpenSearch). Amazon gets the hosting fees, the dependency relationship, and a say in the standard for agentic commerce. Retailers get an AI assistant that knows their catalog, understands their brand voice, and processes customer requests.

Most retailers lack the engineering resources to compete with Amazon. The 60-day deployment timeline (which sounds ambitious) makes this classic “build vs. buy” decision even easier.

Amazon’s own assistant drove nearly $12 billion in incremental sales last year across 300 million users. ASA will succeed because it’s Amazon. The moat is Amazon’s extraordinary network effect; the tech is almost beside the point. Almost.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post Amazon Starts Selling AI Shopping to its Competitors originally appeared here on Shelly Palmer

For Geeks: The exploit is one character. An attacker injects a value into the HTTP Host header. Starlette accepts it, reconstructs the requested URL from it, and the path-based authorization layer waves the request through. The downstream consequences include SSRF, credential theft, and (in some cases) remote code execution. The fix shipped Friday in Starlette 1.0.1.

For Normal People: AI agents work by holding a giant keyring. The keyring opens your email, your calendar, your customer records, your file storage, and every other system you wired the agent into so it could do work for you. The bug infests where the keyrings are stored. Anyone who knows the trick can walk in, grab a keyring, and use every key on it the way your agent would. The damage stops only where the agent’s reach stops.

The official severity (CVSS) rating is 7 out of 10, but X41 D-Sec says that “materially understates” the threat. A one-character header injection that defeats authentication on most production AI tooling without a properly configured firewall earns a 9 or 10 in the real world.

The race to deploy agents has outpaced the security model around them, which puts all of us smack in the middle of an enduring AI arms race. Adjust your enterprise deployment budgets accordingly.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.

The post One Character Broke Millions of AI Agents originally appeared here on Shelly Palmer