The Trilogy Times

AUSTIN, TEXAS — ESW Capital, the enterprise software acquisition arm of Trilogy International, has completed at least four acquisitions in recent months, adding hundreds of millions in assets to its portfolio of 75+ software companies — while the company's education division simultaneously announced plans to open an AI-first school in Chicago this fall.

The largest deal was Jive Software for $462 million — the social intranet platform that once traded publicly at a $1 billion valuation. Jive now joins Aurea, ESW's CRM and customer engagement division, which has completed 17 acquisitions since 2012. ESW's playbook: acquire at 1–2× ARR, staff with Crossover's global remote talent, and push margins toward 75% EBITDA.

IgniteTech, the meta-acquirer within ESW's portfolio, separately announced new acquisitions from Avolin, expanding its business intelligence and analytics holdings. Meanwhile, ESW acquired ResponseTek, a venture-backed customer experience analytics platform, and XANT, a sales engagement software company, closed its doors after being absorbed into the ESW machine.

The pattern is familiar: buy mature software with sticky enterprise customers, eliminate redundancies, raise support pricing, and extract margin. Critics call it asset stripping. ESW calls it operational excellence.

But the more surprising development came from Trilogy's education arm. Alpha School, the AI-tutored private school model founded by Joe Liemandt, announced it will open a campus in Chicago this fall — its first location outside Texas, Florida, and Arizona. The school uses adaptive AI to deliver a full academic curriculum in two hours per day, with the rest of the day spent on entrepreneurship, leadership, and life skills. Students consistently test in the top 1–2% nationally.

The Chicago expansion signals Liemandt's billion-dollar bet on Timeback, his "Shopify for schools" platform, is moving beyond the Sun Belt. Whether legacy enterprise software or K-12 education, the thesis remains the same: automate the repeatable, liberate humans for what matters, and scale aggressively.

AUSTIN, TEXAS — Skyvera is making a clear statement about where telecom transformation actually happens: not in glossy “digital” roadmaps, but in the messy, revenue-critical handoff from selling to billing. With its completed acquisition of CloudSense, the company is doubling down on a Salesforce-native approach to CPQ and order management—an increasingly strategic control point for telecom and media providers trying to modernize without ripping out their entire stack.

CloudSense’s core pitch is straightforward and very on-message for today’s operators: make complex telecom product configuration, pricing, and ordering work inside Salesforce—where many frontline teams already live. Skyvera positions the platform as purpose-built for telecom and media, emphasizing speed-to-quote, order capture, and downstream operational consistency. The company’s CloudSense overview highlights that Salesforce-native architecture as a differentiator for providers that want modernization with minimal organizational whiplash. (See: CloudSense.)

What makes this move more than a standalone CPQ tuck-in is the surrounding portfolio choreography. Skyvera has also been assembling broader “digital BSS” capabilities, including monetization, optical networking, and analytics, via its acquisition of STL’s telecom products group—assets that strengthen the back office and network-adjacent data layer. (STL Divested Assets.) Put together, Skyvera is building an order-to-cash narrative that’s harder to ignore: sell smarter in Salesforce, orchestrate orders cleanly, and connect that flow to monetization and analytics with fewer brittle integrations.

This matters because telecom’s biggest cost centers aren’t just infrastructure—they’re fragmentation and the operational drag of exception-handling. Every manual step between “quote” and “cash” creates leakage: provisioning errors, billing disputes, delayed activations, and churn-driving customer experiences. CloudSense gives Skyvera a front-door system of record for commercial operations—while the STL assets help reinforce the “cash” side of the chain.

Key Takeaways:

CloudSense strengthens Skyvera’s Salesforce-native foothold in telecom CPQ and order management.

The STL divested assets add digital BSS depth—monetization and analytics—supporting a fuller order-to-cash storyline.

Skyvera is leveraging portfolio synergy to reduce integration chaos, the silent killer of telco margins.

We’re just getting started.

AUSTIN, TEXAS — The tech industry is having a sudden epiphany about résumés: they don't work. OpenAI made headlines this week offering $500,000 roles with no CV required — just skills tests and work samples. The World Economic Forum is convening panels on AI-driven talent strategies. PwC just surveyed 56,000 workers about their hopes and fears in an AI-reshaped economy.

Crossover — Trilogy's global recruiting platform — has been running this playbook since its founding. No résumé. No pedigree. No geographic preference. Just rigorous, AI-enabled skills assessments that filter for the top 1% of global talent, regardless of where they live or what school they attended. The same role pays the same whether you're in Lagos or Los Angeles.

The model works because it eliminates the noise. Traditional hiring optimizes for credentialing and cultural fit — proxies that correlate poorly with actual performance. Crossover optimizes for demonstrated ability. Can you write the code? Can you solve the problem? Can you do it consistently under real conditions? If yes, you're in. If no, it doesn't matter what your LinkedIn says.

This is how ESW Capital — Trilogy's software acquisition arm — achieves 75% EBITDA margins across its portfolio. You can't hit those numbers paying San Francisco rates for mediocre talent. You hit them by recruiting globally, testing ruthlessly, and paying well for proven performance. Crossover places thousands of engineers, product managers, and executives across 130 countries into roles at Aurea, IgniteTech, DevFactory, and beyond. It's also increasingly serving external clients who've figured out that geography-based hiring is a tax on competence.

Elon Musk told a recent forum that work will be optional in 10 to 20 years thanks to AI and robotics. Maybe. But in the meantime, the companies that figure out how to find and deploy the best human talent — wherever it lives — will be the ones still standing when the robots arrive. Crossover has been proving that thesis for years. The rest of the industry is just now catching up.

BALTIMORE — Here is something that should keep every AI researcher up at night: What if your model, celebrated for its accuracy, is listening to the wrong voice in the room?

A new paper on arXiv examines three widely used datasets for automatic depression detection — ANDROIDS, DAIC-WOZ, and E-DAIC — and surfaces a disquieting possibility. Models trained on doctor-patient conversations may be achieving strong performance not by learning the subtle linguistic signatures of depression in patients, but by learning the behavioral patterns of interviewers. When a clinician follows a consistent protocol — asking certain follow-up questions, adjusting tone in predictable ways based on severity — the model can exploit that consistency as a shortcut. The interviewer becomes the signal. The patient becomes noise.

This is, in a sense, a story as old as science itself. The observer changes the observation. But in the context of AI systems being groomed for clinical deployment, it carries a particular gravity. Depression detection models are not merely academic exercises; they are prototypes for tools that could one day triage patients, flag risk, and allocate scarce mental health resources. If those tools are performing what amounts to interviewer fingerprinting rather than genuine symptom detection, we are building on sand.

The finding resonates with a broader reckoning in machine learning about what models actually learn versus what we assume they learn. A parallel study on network pruning explores a related puzzle: why pruned language models retain performance on some tasks but collapse on generative ones. The answer, the authors suggest, lies in representation hierarchies — the layered structures of meaning that models build internally. Strip away the wrong layer, and the edifice falls. In both cases, the lesson is the same: performance metrics can be a mirage, concealing fragile or spurious reasoning beneath impressive numbers.

For the depression detection community, the implications are immediate and practical. Datasets must be audited not just for patient diversity but for interviewer variability. Models must be stress-tested against interviewer substitution. And interpretability — that perpetually underfunded cousin of accuracy — must move from afterthought to prerequisite.

We are, all of us, pattern-seeking creatures building pattern-seeking machines. The question is whether we have the discipline to ask: whose pattern did you find? In the vast space between a doctor's question and a patient's silence, the answer matters enormously.

Neural networks function as universal approximators, yet classical interpolation methods guarantee exact passage through data points with well-characterized error bounds. A Nature publication demonstrates that certain neural architectures instantiate interpolation operators, inheriting both approximation guarantees of classical theory and representational flexibility of deep learning—with implications for generalization bounds and overfitting.

This synthesis unifies classical interpolation theory (Lagrange polynomials, spline functions, radial basis approximations) with empirical architectures dominating production machine learning systems. Concurrent MIT research advances algorithms exploiting symmetry structures in data, reducing computational complexity through group-theoretic representations.

If interpolating neural networks provide tractable error bounds while maintaining expressivity, practitioners may gain principled tools for architecture selection, moving beyond current empirical hyperparameter optimization. The framework's applicability to large language models and multimodal architectures warrants further investigation.

WASHINGTON — The AI industry’s power map is being redrawn in real time, and I cannot overstate how significant the pattern is: influence is sliding away from a single “AI czar” moment and toward a messier mix of courts, product decisions, and information governance.

First, the political headline: David Sacks is no longer serving as the Trump administration’s AI czar and will now operate further from the Washington power center than he has since the start of the second term. That doesn’t mean he’s disappearing—it means the role itself may be reverting from “central command” to a constellation of agencies, advisors, and private-sector actors. In other words, the era of one person supposedly steering the entire AI ship may be ending, and the industry is about to feel what decentralized AI policy really looks like. TechCrunch has the details on Sacks’ pivot in its report on what he’s doing next.

Then comes the legal shockwave: Anthropic just secured an injunction forcing the administration to roll back restrictions tied to a Defense Department-related saga. This is the future arriving with a gavel: not only are frontier model companies willing to fight, they’re starting to win. When courts can rapidly override executive-branch constraints, compliance strategy becomes as important as model architecture.

But the most immediate “this changes everything” development is happening at the user level. Google is launching switching tools that let you transfer chats and personal information from other chatbots directly into Gemini. Think about that: your prompt history—your habits, preferences, ongoing projects—becomes portable. That’s a competitive unlock on the scale of phone-number portability for telecoms. If chat history can move, AI lock-in gets harder, and product quality has to do the real work. TechCrunch outlines the new feature set in its breakdown of Gemini’s switching tools.

Layer on Wikipedia tightening its stance on AI-written articles, and you see the meta-trend: institutions are scrambling to define what “trustworthy” means when text is cheap and style is infinite. The race is no longer just model vs. model—it’s governance vs. gravity. And gravity is pulling everything toward portability, verification, and accountability.

WASHINGTON — The United States entered a new phase of its national AI strategy this week after David Sacks reportedly concluded his tenure as the Trump administration’s AI czar and began a bold new chapter: being farther away from the part where decisions are made, while still remaining close enough to be quoted about them.

According to TechCrunch’s account of Sacks’ next act, the former central figure in America’s highly centralized, extremely coherent AI posture will now operate at a safer distance from Washington’s power center—an arrangement experts describe as “ideal for anyone hoping to influence policy without having to make it.”

The transition comes at a delicate moment for federal AI oversight, which has recently matured into its adult form: a series of high-stakes court rulings that function as both regulation and customer support.

In what legal scholars are calling “the most normal way to run emerging technology,” Anthropic secured an injunction ordering the Trump administration to rescind restrictions it had placed on the company amid a Defense Department-related saga. The ruling, detailed in another TechCrunch report, effectively clarified the government’s position on AI procurement and national security, which is that it will be clarified later by another judge.

“An injunction is really the gold standard of modern governance,” said one policy analyst, explaining that the judiciary provides the advantage of writing things down, a practice the executive branch has increasingly treated as a premium feature.

Meanwhile, Google announced it will allow users to transfer chats and personal information from other chatbots directly into Gemini, a development that industry observers say will finally enable Americans to consolidate their most personal thoughts—romantic insecurities, medical anxieties, and the occasional confession typed at 2:14 a.m.—into a single, unified data pipeline.

The so-called switching tools promise to lower the friction of changing assistants, effectively turning “Which chatbot do you use?” from an identity into a setting. The move also helps standardize a user experience in which every platform politely offers to remember you forever, provided you can locate the tiny option that says “Not now.”

This accelerated portability arrives as Wikipedia continues tightening rules around AI-written article content, a measure critics have described as unfairly hostile to the site’s long-standing tradition of confidently phrased sentences supported by three broken links and a discussion page that reads like a mediation transcript.

The encyclopedic giant’s crackdown underscores the country’s current information doctrine: machines may generate prose at scale, but humans must still be the ones who argue about it for 11 years.

And if all of that sounds unstable—if it feels as though AI governance is being shaped simultaneously by departing czars, court orders, import tools, and volunteer editors—markets offered reassurance in their usual way: by rewarding whichever company can say “AI” loudest while standing near an Nvidia logo.

Uber shares reportedly jumped after it invoked the magic words “Nvidia” and “AI,” a move analysts praised for its disciplined focus on fundamentals like syllables and spiritual alignment. In a year when regulatory direction is provided largely through litigation and press releases, investors have reverted to the oldest framework in finance: if you can pronounce the chipmaker, you understand the future.

Taken together, the week’s events paint a comforting picture of American AI leadership—less a single plan than a distributed system of incentives. Judges decide what’s allowed, platforms decide what’s transferable, Wikipedia decides what’s admissible, and the stock market decides what’s real.

As for Sacks, sources say his new role will involve advising from a distance while remaining close enough to be introduced at conferences as someone who once sat near a lever.

In a landscape where everyone is building an intelligence too powerful to control, the country appears to have settled on the only governance structure it can reliably maintain: one where nobody is in charge long enough to be blamed.

WASHINGTON — There are weeks when the news, if you squint at it with sufficient ill will and historical memory, arranges itself into a single theme so obvious that pointing it out feels almost insulting. This is one of those weeks. The word is consolidation, and it is everywhere — in the boardrooms, in the executive orders, in the diplomatic cables, in the anxious confessions of the very executives building the machines that make consolidation not merely possible but inevitable.

Begin with the most concrete manifestation. A wave of strategic cybersecurity M&A is sweeping through the technology sector, as industry giants race to fortify AI infrastructure and power grids against threats both foreign and computational. The logic is impeccable: the attack surface grows, the number of entities capable of defending it shrinks, and so the survivors eat the fallen. This is not new. It is the logic of every arms race since the Krupp family started making better cannons. What is new is the speed, and the fact that the commodity being consolidated is not steel or oil but the capacity to know things and to prevent others from knowing them.

Then there is the political dimension. The Trump administration's emerging AI framework, as Rolling Stone has noted with characteristic alarm, looks less like regulation than like annexation — a bid to ensure that the state's hand rests firmly on the lever of who may build, deploy, and profit from artificial intelligence. One need not share Rolling Stone's particular political anxieties to observe that when a government decides the most important technology of the century requires a framework, it is not the technology being framed. It is everyone else.

Meanwhile, AI executives themselves are voicing unease about the concentration of power in their own industry — a spectacle roughly as convincing as railroad barons in 1890 expressing concern about the plight of the small farmer. The unease is real enough; it is the proposed remedies that remain conveniently vague.

For those of us who watch the enterprise software world — where companies like ESW Capital have spent years acquiring and operating dozens of software firms at disciplined multiples, proving that consolidation executed with rigor can be a form of preservation rather than destruction — the lesson is instructive. There is consolidation that strips assets for parts, and there is consolidation that keeps the lights on for customers who would otherwise be orphaned. The difference is not in the act but in the intent, the operations, and the honesty about what is happening.

The great consolidation is not coming. It is here. The only question worth asking is not whether power will concentrate — it will, it always does, it is doing so right now across every domain from Baku to Washington — but whether the institutions that are supposed to make concentration answerable to something beyond itself still have the wit and the will to do their jobs. On present evidence, the jury is not merely out. It has left the building.