There is a category of production incident that engineering teams are not tracking yet — because it doesn’t fit any existing postmortem template.
The agent initiated an action. The action was technically correct given the agent’s context. The context was incomplete. The infrastructure cascaded. And, by the time the incident review happened, three teams were arguing about whether it was an agent failure or an infrastructure failure, because the frameworks for thinking about these two things have never been connected.
The scale of this exposure is no longer theoretical. Seventy-nine percent of organizations now have some form of AI agent in production, with 96% planning expansion. Gartner predicts 33% of enterprise software will include agentic AI by 2028, but separately warns that 40% of those projects will be canceled due to poor risk controls.
What neither statistic captures is the failure mode happening between those two numbers: Agents that are running, that are not canceled, and that are quietly generating infrastructure events no one has categorized as risk.
Advertisement
I’ve spent six years building infrastructure automation systems at enterprise scale, first at Cisco (leading AI-driven lifecycle platforms deployed across 20-plus global enterprise customers), then at Splunk (designing AI-assisted root cause analysis and observability workflows across thousands of enterprise environments).
During that time I also filed a patent on intent-based chaos engineering methodology. And across all of it, I kept watching organizations make the same structural mistake: Treating autonomous agents and chaos engineering as separate disciplines. They are not. They are the same discipline, and the gap between them is quietly generating the next wave of major production incidents.
The judgment call that agents skip
To understand why this matters, you need to understand what’s actually broken in how enterprises govern chaos today, before you add agents to the picture.
Most mature engineering organizations have invested in chaos engineering programs. Game days, blast radius controls, SLO-gated experiments. When a human engineer initiates a chaos experiment, the sequence has a critical property: A human is making a judgment call about whether the system has capacity to absorb the perturbation right now. They check dashboards. They look at the error budget burn rate. They assess whether dependencies are stable. It’s imperfect and often intuitive, but there is at least a person in the loop asking the right question before anything runs.
Advertisement
When you introduce an autonomous remediation agent, one that can restart services, reroute traffic, scale resources, or modify configurations in response to detected anomalies, that question disappears. The agent sees an anomaly. The agent takes an action. The action is a chaos event. No SLO burn rate check. No blast radius calculation. No human judgment about whether right now is the right moment to introduce additional stress into a system that may already be under pressure from three other directions.
Here is the specific failure mode I have watched play out. A remediation agent detects elevated latency on a microservice and responds by restarting the service cluster; a reasonable action given its training data and its narrow view of the incident. What the agent doesn’t know: Three other services are in the middle of handling peak traffic. The shared connection pool is already at 87% utilization. A dependent database is running a background index rebuild. The restart triggers a thundering herd against the recovering service.
What started as a latency spike the agent was designed to fix becomes a cascade the agent was never designed to model. The blast radius of that agent action was not the service restart. It was everything downstream of the restart, in a system state the agent had no complete picture of.
Nobody’s chaos engineering program had tested for that specific combination. Nobody’s blast radius calculation had included the agent as an actor. Because we don’t think of agents as chaos injectors. We should.
Advertisement
According to the AI Incidents Database, reported AI-related incidents rose 21% from 2024 to 2025. That count almost certainly understates the actual exposure, because most organizations have no incident classification that captures an autonomous agent action as the initiating cause of a cascade. The incident gets logged as a service restart, a connection pool saturation, or a latency event. The agent is invisible in the postmortem.
Absorb capacity is a resource; most systems don’t treat it that way
The underlying problem is that enterprise systems have no shared language for absorb capacity — the real-time estimate of how much additional stress a system can take before it breaches its SLO commitments. Chaos engineering programs manage it implicitly, through human judgment and static thresholds that fire after a limit has already been crossed. Agents don’t manage it at all.
Through structured primary research with site reliability engineering (SRE) and platform engineering practitioners across organizations including Intuit and GPTZero, I’ve been developing a resilience budget model. The core idea is to treat absorb capacity as a continuously recomputed, consumable resource rather than a static threshold you try not to breach.
A resilience budget draws on four live signal classes.
Advertisement
SLO burn rate is the primary input, because it directly encodes the distance between current system behavior and the commitment that actually matters. If a system is burning its monthly error budget at five times the expected rate, the resilience budget is near zero regardless of what CPU utilization looks like.
P99 latency trend matters more than absolute latency, because a service trending upward over forty minutes tells you something different than a service that has been stable at the same absolute value.
Dependency saturation state is the most commonly missed signal; a chaos experiment or an agent action that assumes a shared connection pool is freely available when it’s sitting at 87% will produce failure modes that nobody designed for.
Application behavioral signals, session completion rates, API call pattern shifts, conversion degradation, and surface system stress earlier than infrastructure metrics do, because users feel the degradation before Prometheus reports it.
What makes this a budget rather than a threshold is that it is consumable. Every chaos experiment draws from the available capacity. Every agent action draws from it. In multi-team organizations where multiple experiments and multiple agents may be acting simultaneously, the budget is shared.
Without a shared ledger of consumption, two teams running experiments against overlapping dependencies produce a combined blast radius that neither team planned. Add autonomous agents acting completely outside the ledger, and the accounting collapses.
Image provided by author.
Where language models help, and exactly where they fail
Several engineering organizations are now running experiments using large language models (LLMs) to generate chaos hypotheses from dependency graphs and incident postmortem corpora. The results are directionally useful. Language models surface plausible failure modes that experienced SREs recognize as worth testing, and they generate hypotheses faster than manual processes, particularly when working from rich postmortem history.
Advertisement
The limit is dependency graph staleness, and it is a hard limit. A hypothesis generated from a graph that doesn’t reflect last month’s service extraction, or a new shared library dependency added two sprints ago, will propose an experiment with incorrect blast radius assumptions. The problem is not that the model makes a mistake, it’s that the model doesn’t know it’s making one. It will be confidently incorrect about a system boundary that no longer exists, and in chaos engineering, confident incorrectness in production means an unplanned outage.
Stanford’s Trustworthy AI Research Lab found that model-level guardrails alone are insufficient: Fine-tuning attacks bypassed leading models in the majority of tested cases. The implication for chaos hypothesis generation is direct, a model that cannot reliably hold its own safety boundaries cannot be trusted to accurately model the blast radius of an action it has never seen in a dependency graph it has not verified.
When hypothesis generation draws instead from postmortem corpora, the staleness problem shrinks considerably. Postmortems describe failures that actually occurred in the system at a specific moment in time. The signal is inherently validated by production reality. This is the tractable near-term AI application in this space, and it is genuinely useful for organizations with mature incident documentation practices.
What AI cannot do, and should not be asked to do, is make the execution decision when signals are ambiguous. That judgment requires awareness of things that live entirely outside any monitoring system: Pending deployments that changed the dependency landscape an hour ago, on-call staffing levels on a holiday weekend, a customer commitment that makes any additional risk unacceptable until Monday.
Advertisement
A model without access to that context should not be making that call. This is not a temporary limitation pending a more capable model. It is a structural constraint of what machine observability can represent, and building an agent architecture that ignores it is building one that will eventually make a consequential decision with incomplete information — and no human in the loop to catch it.
What this means for how enterprises govern agents in production
The governance implication is straightforward to describe and harder to implement than it sounds. Every autonomous agent action that touches infrastructure needs to register against the same live signal layer that governs chaos experiments. The same SLO burn rates, latency trends, dependency saturation states that a human engineer would check before initiating an experiment should gate what an agent is permitted to do and when. If the resilience budget is below a defined floor, the agent waits or escalates. It does not act.
Agent actions also need to be modeled as experiments, not just logged as events. When an agent restarts a service, the question isn’t only whether the restart completed successfully. It’s whether the blast radius of that action was proportionate to the available absorb capacity, and what cascading effects it produced across dependencies. That is chaos engineering data. It belongs in the budget model, feeding the next decision the agent or the team needs to make.
And when signals are genuinely ambiguous, when the budget score is unclear, when a recent deployment has changed the topology in ways the agent’s context window doesn’t capture, when dependency states are in flux, the execution decision needs to go to a human. Not as a permanent limitation on agent autonomy, but as a hard engineering requirement for the current state of the technology.
Advertisement
A circuit breaker that hands ambiguous cases to a human is not a weakness in the agent architecture. It is the thing that makes the architecture trustworthy enough to actually run in production. Intent-based verification formalizes exactly this: Defining what correct agent behavior looks like before deployment, then continuously probing whether those boundaries hold under live system conditions.
The organizations that operate autonomous agents reliably at scale are not the ones with the most sophisticated models. They are the ones that understood, before something went badly wrong, that every agent action is a chaos event and built their governance layer accordingly.
The practical first step is unglamorous: Audit every autonomous agent currently touching infrastructure, map its action surface against your live SLO burn rate signals, and define explicit floor conditions below which the agent is required to wait or escalate. That audit will surface agents acting entirely outside your resilience accounting.
Most organizations running agents at scale today have several. Find them before production does.
Advertisement
Sayali Patil has spent 6-plus years at Cisco Systems and Splunk building the reliability and automation systems that keep enterprise AI infrastructure running at scale.
Welcome to the VentureBeat community!
Our guest posting program is where technical experts share insights and provide neutral, non-vested deep dives on AI, data infrastructure, cybersecurity and other cutting-edge technologies shaping the future of enterprise.
Read more from our guest post program — and check out our guidelines if you’re interested in contributing an article of your own!
The cover of Newsweek magazine, May 20, 1996 — exactly 30 years ago today.
Take a breath, close your eyes, and think about the words that define Seattle.
Innovative. Outdoorsy. Global. Inventive. Smart. Progressive. Independent. A little reserved. A little weird.
Thirty years ago today, Newsweek magazine published a cover story featuring political journalist Michael Kinsley titled: “Swimming to Seattle: Everybody Else Is Moving There. Should You?”
Back in May 1996, Seattle was emerging as one of America’s great boomtowns: grunge, coffee, software, airplanes, the web. A place with talent, ideas, ambition and room to grow.
Advertisement
It’s one of the reasons why I moved here 30 years ago, from a small town in Ohio.
Today, Seattle remains one of the world’s most important innovation hubs, home to global technology giants, leading AI research, world-class research and extraordinary entrepreneurial talent.
Which is exactly why the city’s shifting national image should concern us.
Because a new narrative about Seattle is taking hold nationally. And unlike the rain-slicker caricatures of the 1990s, this one isn’t charming.
Advertisement
The emerging narrative is this: Seattle has become increasingly ambivalent — even hostile — toward the very industries and innovators that helped build its prosperity.
And it’s not just the national media. Seattle’s KOMO News reported this week on remarks by former Washington state governor Chris Gregoire, who pointed out a ballooning state budget since she left office in 2013.
“I would suggest to you, we don’t really have an income problem, we have a spending problem,” Gregoire said at a meeting hosted by the Association of Washington Business earlier this month.
You may disagree with those headlines. You may dislike the politics behind them. But rhetoric, image and storytelling matter — especially in a moment when cities are competing fiercely for talent, investment, startups and relevance in the AI era.
Advertisement
And right now, Seattle’s story is drifting in the wrong direction.
This week, the chairman of an iconic Seattle company — not operating in the tech industry — told me that the city’s increasingly anti-business image was complicating a national CEO search. Meanwhile, entrepreneurs and investors regularly tell us they feel vilified or unwanted.
We’ve spent more than 50 years importing some of the smartest people on the planet to this corner of the world — people working on things like cancer research, robotics, and yes AI — only to turn around and tell them not to let the door hit them on the way out.
Cities compete on psychology as much as policy. And our psychology is a bit shattered right now.
Advertisement
Six years ago, another national narrative engulfed Seattle during the Capitol Hill Autonomous Zone, or CHAZ — a protest occupation in Seattle’s Capitol Hill neighborhood that formed during the 2020 national reckoning over policing and racial justice.
Living here at the time, I thought much of the national media portrayal was exaggerated. I remember assuring friends and family back in Ohio that Seattle had not, in fact, descended into dystopian chaos despite what cable news suggested.
This moment feels different.
The concern now isn’t lawlessness or political theater. It’s civic drift. And right now, the national headlines resonate. They are telling a real story.
Advertisement
Seattle’s uncertainty about the very economic engine that transformed it into a global city is something that competitors are already beginning to notice.
Contrast Seattle with San Francisco, another progressive West Coast city wrestling with many of the same challenges. Its leaders are aggressively selling a comeback narrative centered on AI, entrepreneurship and reinvention.
Seattle, by comparison, is a city arguing with its own success.
Of course, there has always been a strain of “Lesser Seattle” thinking woven into Seattle’s culture — the instinct to resist growth, keep outsiders away and preserve an earlier version of the city before construction cranes and rapid change arrived.
That sentiment isn’t entirely irrational. Growth brought real costs: affordability challenges, displacement, congestion, inequality.
But it also brought extraordinary opportunities.
Advertisement
And in an era when artificial intelligence is reshaping industries, cities cannot afford to become complacent, confused about their identity, or dismissive of the people and companies driving innovation.
Seattle still has remarkable advantages. But advantages are not permanent.
Cities rise because they project confidence, ambition, and possibility. They decline when they begin treating success as something inevitable — or worse, something suspect.
Maybe that’s why another piece of Seattle culture has been stuck in my head lately: the absurdly catchy 1996 song “Peaches” by the Seattle rock band The Presidents of the United States of America: “I’m movin’ to the country, I’m gonna eat me a lot of peaches.”
Advertisement
The song captured a certain quirky, ironic version of Seattle at the tail end of the grunge era, a city that didn’t take itself too seriously.
Right now, though, Seattle faces a much more serious question: What kind of city does it actually want to become?
The choice seems clear. Move forward, progress, and tell a fresh story of hope in a city that’s still swimming in opportunity.
Ready to portal-jump back into the Morty-verse for Rick and Morty season 9? Perfect – we’ll show you how to watch season 9 from anywhere when episode 1 drops on May 24 in the US. Brits won’t have to wait long this year – season 9 drops on May 25 thanks to HBO Max.
The madcap irreverence that earned Rick and Morty such a huge and loyal following in the first place is still very much the backbone of the show, as demonstrated by Beth and Summer’s altercation with anthropomorphic furniture, the floor turning to lava and the plentiful pop culture riffs, which for the latest instalment include Planet of the Apes and Kill Bill (Pai Mei unleashes the Five-Point-Palm Exploding-Heart-Technique on Rick).
Bewildered? Scroll down to watch the trailer and achieve total clarity… or at least become ever so slightly less perplexed.
Advertisement
Fans hoping to see more of an overarching storyline may be disappointed, although there are signs that poor Morty’s deeply held sadness and exasperation are building towards something. There are introspective elements to some of this season’s sci-fi adventures, and they’re unlikely to reflect too well on Rick.
Read on as we explain how to watch Rick and Morty season 9 from anywhere and potentially for free.
Can I watch Rick and Morty season 9 for free?
Yes. YouTube TV carries Adult Swim as standard, and offers new subscribers a whopping 21-day FREE trial.
That means you can watch at least the first few episodes of Rick and Morty S9 without charge.
Advertisement
Traveling abroad right now? You can use a VPN to access your YouTube TV account as if you were right at home.
Use a VPN to watch Rick and Morty season 9 from anywhere
A VPN is a handy piece of software that can make your device appear as if it’s back in your home country and unlock your usual streaming services. The best VPN right now? We recommend NordVPN – it does everything and comes with up to 75% off.
Advertisement
How to watch Rick and Morty season 9 in the US
Rick and Morty season 9 premieres on Adult Swim at 11pm ET/PT on Sunday, May 24. Further episodes will go out one at a time in the same slot weekly.
If you don’t have the channel on cable, it’s available through YouTube TV, which normally costs $82.99/month but comes with a 21-day FREE trial and $15 off each of your first five months.
Have one of these subscriptions but away when Rick and Morty S9 is on? You can still access your usual streaming services from anywhere by using a VPN.
Advertisement
How to watch Rick and Morty season 9 in the UK
If you’re used to Rick and Morty episodes going out on free-to-air E4 in the UK, it’s all change for season 9, which is – at least initially – exclusive to HBO Max.
New episodes become available every Monday, starting May 25.
HBO Max prices starts at £5.99/month, though it comes bundled with all Sky TV plans.
Advertisement
At the time of writing it isn’t clear if Rick and Morty season 9 will also air on E4. S8 landed a week after its US premiere.
Away from home? You can still connect to your usual streaming services by downloading a VPN and pointing your location back to the UK.
How to watch Rick and Morty season 9 in Canada
In Canada, Rick and Morty season 9 airs on Adult Swim, going out at 11pm ET/PT on Sundays, starting May 24.
Advertisement
Cord cutters can tune in via Stack TV, which carries a number of channels, including Adult Swim, Global TV and National Geographic, and is available via Prime Video, Fubo and more. It’s free to Prime Video subscribers for the first seven days and CA$14.99/month thereafter.
Away from Canada now? Use NordVPN to watch your usual streaming service when overseas.
How to watch Rick and Morty season 9 in Australia
Rick and Morty season 9 premieres on Monday, May 25 on HBO Max in Australia.
Advertisement
A subscription starts at AU$11.99/month.
Outside Australia? Aussies away from home can use a VPN to unblock HBO Max and watch Rick and Morty S9as they would at home.
Rick and Morty season 9 – Need to Know
Advertisement
Rick and Morty season 9 trailer
Rick and Morty | Season 9 Official Trailer | adult swim – YouTube
Season 9 of Rick and Morty premieres in the US and Canada on Sunday, May 24 and continues weekly for 10 episodes.
In the UK and Australia, episodes will arrive a day later, starting from Monday, May 25.
Rick and Morty S9 episode guide
Season 9 of Rick and Morty comprises 10 episodes, which will air in North America on the following schedule:
Episode 1 – Sunday, June 24
Episode 2 – Sunday, June 31
Episode 3 – Sunday, July 7
Episode 4 – Sunday, July 14
Episode 5 – Sunday, June 21
Episode 6 – Sunday, June 28
Episode 7 – Sunday, July 5
Episode 8 – Sunday, July 12
Episode 9 – Sunday, July 19
Episode 10 – Sunday, July 26
We test and review VPN services in the context of legal recreational uses. For example: 1. Accessing a service from another country (subject to the terms and conditions of that service). 2. Protecting your online security and strengthening your online privacy when abroad. We do not support or condone the illegal or malicious use of VPN services. Consuming pirated content that is paid-for is neither endorsed nor approved by Future Publishing.
Getting authentic smoky flavour from a backyard grill usually means committing to a bulky charcoal setup, hours of babysitting the temperature, and a clean-up job that extends well into the evening.
The woodfire flavour here comes from real hardwood pellets, not gas or charcoal, and the integrated smoke box circulates that smoke evenly around the food, so the result tastes genuinely earned rather than artificially added.
Advertisement
Beyond smoking, the grill operates across seven functions, including grilling, air frying, roasting, baking, broiling, and dehydrating, which means it covers the full spread of outdoor cooking without requiring a separate appliance for each method.
Two built-in thermometers let you monitor two different proteins simultaneously and set individual doneness targets for each, which is the kind of practical feature that stops a cook from having to hover anxiously over the grill for the duration of the meal.
The Ninja ProConnect app connects via Wi-Fi and Bluetooth, sending real-time notifications to your phone when the grill is preheated, when to flip, and when your food is ready, so you can actually be present with your guests rather than tethered to the cooking area.
Advertisement
With 180 square inches of cooking space, the Ninja Woodfire Pro Connect handles up to 10 burgers, two full racks of ribs, or a 12-pound brisket in a single cook, which is more than enough capacity for a proper summer gathering.
The stainless steel construction is rated weather-resistant for year-round outdoor storage, meaning it does not need to be hauled inside between uses, and the unit runs on electricity rather than gas, so there is no cylinder to manage or replace.
This is genuinely a strong buy for anyone who wants authentic smoky results without committing to the complexity of traditional BBQ, and a 20% discount heading into summer makes the case even harder to dismiss.
I almost don’t want to call Meshchera a match-three game because I fear that kind of undersells how captivating it is. But, it is a game you play within a six by six grid, in which you have to group matching tiles in clusters of three or more so they may merge and become other, higher value tiles, so that’s the description we’re working with. The atmosphere is off the charts, though, which isn’t something I’m used to finding in these types of games. It has gorgeously detailed artwork and background music that you can get completely lost in.
In Meshchera, you can choose to go for the high score or pick from several challenges that will dictate how you approach the round, like “kill five monsters” or “keep 10 monsters for 10 turns.” The gameboard is a dark marsh that will slowly become overrun with vegetation and creatures, unless you can stay ahead of creep by skillfully matching tiles to condense them into other things. Grasses become flowers, which become trees, campfires, houses, churches, etc. It is a uniquely complex matching game — you’re given next to no information about how the items work or how different elements on the board behave and interact, so you have to figure it out along the way and course-correct as you learn.
khvoshch
I’ve spent quite a bit of time playing Meshchera over the last week, but certain things still elude me. Take the “create and destroy a Monster” challenge. I have absolutely no idea how to create a monster, and that’s not for lack of trying. But, this gives me something to keep working toward even as my high scores nudge higher and higher. The game includes 10 challenges right now, and the developer says more are coming soon. Meshchera is really good, and feels like the kind of game you can revisit ad infinitum. It’s already found itself a home in my folder of “go-to” Playdate games.
Meshchera isn’t available in the Playdate Catalog (yet?), but don’t let that stop you from trying it out. It’s on itch.io at the moment, and sideloading games onto the Playdate is incredibly easy. Once you have the game file, you can just drag and drop it right into your library by signing into your Playdate account and going to the Sideload tab. This can also be done via USB. Panic has a detailed explanation of all the options, if you need some guidance.
Twelve AI labs have a combined valuation larger than Ford and GM. None of them sell anything. I call them the Virgin Unicorns — valued above a billion dollars, but innocent of product or revenue.
OpenAI proved that an AI research lab with the right product could become one of the most valuable companies on earth. A dozen other AI labs are trying to repeat the trick. They have raised more than $29 billion at a combined valuation approaching $130 billion, without shipping anything a customer can buy.
Two questions are worth asking:
Why are sophisticated investors writing growth-stage checks to pre-companies?
* Limited research release. Tinker is a fine-tuning tool for researchers; Marble is a 3D-world-generation API in early partner access. Neither is a general-availability commercial product. Sources: company announcements, Bloomberg, Financial Times, TechCrunch, Crunchbase, and PitchBook reporting from 2024-2026. Valuations reflect the most recent confirmed round; figures for rounds in active negotiation are not included.
To answer these questions, let’s identify four patterns across this cohort of companies.
Pattern 1: The pedigree premium. Every founder is a recognized leader in their field, and most come from a small set of institutions. Roughly four-fifths hold PhDs, mostly in computer science from a handful of universities — Berkeley, Stanford, MIT, Toronto, Alberta, Cambridge, UCL — and most of the rest left PhDs at one of those programs to start their companies.
Advertisement
On the employer side, the concentration is tighter still. Four of the twelve companies are anchored by DeepMind alumni (Ineffable, Reflection, Ricursive, Recursive Superintelligence). Two are anchored by OpenAI alumni (Thinking Machines, Safe Superintelligence). AMI Labs traces back to Meta’s FAIR group, and Humans& draws its founders from across Anthropic, xAI, and Google. Stanford and Berkeley faculty appointments account for most of the rest (World Labs, Physical Intelligence, and Noah Goodman of Humans&).
Four institutions — DeepMind, OpenAI, Berkeley, and Stanford — have produced the founders of nearly every Virgin Unicorn in the table. Investors are pricing CVs, not products.
Pattern 2: Nvidia as kingmaker. Nine of the twelve companies in the table have Nvidia as an investor. The supplier of the picks and shovels is also an equity holder in the prospectors. Nvidia gets early visibility into the most ambitious AI bets, locks in compute commitments, and earns multiples on capital deployed at near-zero marginal cost. Selling the shovels was already a good business. Owning the mines too is unprecedented.
Pattern 3: The cap tables are unusually wide. Each round in the table includes a syndicate of ten to twenty investors — venture firms, corporate strategics, sovereign wealth funds, and individuals. Sequoia and a16z still lead. But the rounds are large enough that they require balance-sheet capital — from JPMorgan, BlackRock, Alphabet, the UK Sovereign AI Fund, Samsung, Temasek, ADIA, and Bezos personally — to fill out. That makes these rounds structurally different from classical venture financings.
Advertisement
Pattern 4: A post-LLM thesis. Every company is arguing, in some form, that the current paradigm isn’t enough — that scaling LLMs won’t reach AGI, and that something else (world models, reinforcement learning, agentic systems, AI scientists, novel chips, formal mathematical reasoning) is required. The thesis is the product. The product is a promise.
Others have dissected these unicorns:
Howard Marks, in his December 2025 Oaktree memo Is It a Bubble?, described investor behavior as “lottery-ticket thinking” — investors backing startups with no product on the dream of an enormous payoff despite an overwhelming probability of failing.
Derek Thompson, writing in October, framed the same dynamic by reporting that a Thinking Machines pitch meeting was described by one investor as “the most absurd pitch meeting” because Mira Murati “couldn’t answer any questions” about what she was building.
GeekWire’s own year-end survey of regional venture investors found the same skepticism closer to home: the bubble, they said, is most pronounced at the early stages, where AI storytelling can substitute for real traction.
The lottery-ticket framing is now conventional wisdom. But will this lottery pay out? One way to handicap the odds is to look to the past.
What history teaches us
The closest historical parallel is not the dot-com era. Webvan, Pets.com, and Boo.com failed not because they were pre-product, but because they had products and bad business models. Those companies burned capital on infrastructure and marketing, not on research.
The closer cautionary tales are the celebrity-founder pre-product flops of the last fifteen years.
Advertisement
Magic Leap raised $3.5 billion over nine years on the strength of Rony Abovitz’s prior exit and shipped a flop.
Quibi raised $1.75 billion on Katzenberg and Whitman’s pedigree and lasted six months.
Inflection AI raised $1.5 billion on Mustafa Suleyman and Reid Hoffman and was effectively absorbed into Microsoft in 2024 — its team hired, its technology licensed, its company hollowed into a shell.
In each case, founder credentials raised the money. The product never justified the valuation.
The structurally closest analogy, though, is biotech. Roughly 80% of 2021 biotech IPOs were pre-revenue. The probability that a pre-clinical drug reaches commercialization is under 10%. Development takes a decade and costs $1 billion. Yet a Bentley University study of 319 biotech IPOs from 1997 to 2016 found that the cohort produced over $100 billion in net shareholder value despite a failure rate above 50%. The winners were large enough to carry the portfolio. And many of the most successful biotechs were acquired before reaching profitability.
The Virgin Unicorns are biotech-shaped businesses. Pre-revenue, science-driven, decade-long timelines, binary outcomes, acquisition as the usual exit. But they aren’t financed like biotechs. Biotech investors release capital in milestone tranches tied to specific scientific results, and they expect most candidates to fail. Virgin Unicorn investors release capital in one large round on the strength of a CV, and price for success. Same shape of business, opposite financing logic. That mismatch is where the disappointment will come from.
Why Sequoia invests anyway
The OpenAI story counters the biotech analogy. From its 2015 founding to the ChatGPT launch in late 2022, OpenAI looked exactly like a Virgin Unicorn — pre-consumer-product for seven years, billions in capital, and only research to show for it. Then ChatGPT shipped and revenue went from zero to over $10 billion in three years. No biotech has ever scaled like that.
Sequoia and other investors writing checks to today’s Virgin Unicorns aren’t pricing for biotech outcomes. They’re pricing for the second coming of OpenAI.
The table above makes the size of that bet legible. Early-stage venture investors aim for a 10x return. Most of these twelve will return zero, so the one winner has to carry the other eleven by itself. At a $127 billion aggregate marked-up value, that means the winner alone has to produce something like $1.3 trillion in value.
That is not a forecast — it is the bet the VCs have already placed. Sequoia and a16z made exactly this kind of bet on OpenAI and Anthropic, and the on-paper returns have already vindicated it many times over. Anthropic itself looked like a Virgin Unicorn in 2022 — and then it shipped Claude and built revenue.
The historical record suggests some skepticism. But bubbles have a way of producing the occasional Amazon or Google amid the wreckage. Identifying which Virgin Unicorn will become a trillion-dollar company — a “kilocorn,” a thousand unicorns in one — is tough. Which one would you bet on?
Not all bad news: Crypto billionaire signs up for a mission to Mars
SpaceX called off the launch of its huge Starship rocket seconds before liftoff due to a ground equipment problem.
The countdown clock reached a planned hold at T-40 seconds after a relatively trouble-free process. Some iffy weather had cleared, and everything looked good for the twelfth Starship test flight – the first try-out for the latest generation of the vehicle and launchpad.
Advertisement
Alas, it was not to be. After repeatedly resetting the countdown clock to the T-40 second mark due to problems, which included warnings from sensors on the quick-disconnect arm on the launch pad and issues with the pad’s water diverter, SpaceX eventually threw in the towel and scrubbed the launch.
Boss man Elon Musk blamed the scrub on the ground equipment, and posted on X: “The hydraulic pin holding the tower arm in place did not retract.” Musk wrote that if the issue could be fixed, SpaceX will try again later today. The next window opens at 5:30 pm CT, according to the billionaire.
Considering that this was the first launch attempt from a new pad and the first of this vehicle’s iteration, the countdown problems are unsurprising. As such, getting to the T-40 second mark was an achievement in its own right. Sadly, the team had only a few minutes to deal with the problems, since the propellant loading was complete and the fuel temperature could not be maintained for long.
Expectations are high for this mission. Despite years of development and Musk’s promises, Starship is still non-operational, and its launches remain on suborbital trajectories during its test phase. The vehicle has quite a way to go before it can play a part in NASA’s goal of landing a crew on the Moon.
Advertisement
According to the company’s recent IPO filing, “We expect Starship to commence payload delivery to orbit in the second half of 2026.”
The second half of 2026 is only weeks away, so it’ll be an interesting few months.
The IPO filing also states that Musk’s performance-based restricted shares in SpaceX vest upon the establishment of a permanent human colony on Mars “with at least one million inhabitants.”
First, however, the SpaceX needs to get to Mars. During the scrubbed launch attempt, it announced that crypto billionaire Chun Wang, who commanded the Fram2 private human spaceflight mission in 2025, would be on the crew for a future flyby of the red planet.
Advertisement
Hopefully, Wang’s jaunt to Mars won’t end up canceled like the dearMoon project, a mission to the Moon financed by Japanese billionaire Yusaku Maezawa. The project was unveiled in 2018, but was eventually canceled in 2024. Starship has yet to hit Earth orbit, let alone head to the Moon. ®
The annual tech showcase highlights next-gen AI, cloud, and future-ready ICT solutions while uniting ecosystem partners to build the foundation for the nation’s AI era
Partner Content ZTE Corporation (0763.HK / 000063.SZ), a global leading provider of integrated information and communication technology solutions, held ZTE Day Indonesia 2026 in Jakarta, as its annual technology showcase event, bringing together industry leaders, technology partners, and digital ecosystem players to discuss the future of AI, intelligent infrastructure, and digital transformation in Indonesia.
ZTE reinforced its commitment to accelerating Indonesia’s digital economy growth through intelligent and future-ready ICT solutions
As industries increasingly adopt AI, cloud technologies, and data-driven operations, the demand for smarter, adaptive, and future-ready digital infrastructure continues to accelerate. Responding to this momentum, ZTE Day Indonesia 2026 highlighted how AI, intelligent networks, cloud infrastructure, and next-generation connectivity are becoming key foundations for national digital competitiveness and future economic growth.
The event showcased a broad range of integrated ICT innovations spanning artificial intelligence (AI), intelligent computing, cloud infrastructure, optical transport, enterprise networking, Wi-Fi 7, and next-generation connectivity technologies designed to support enterprises, operators, and industries navigating the AI era.
Advertisement
Liu Sen, President Director of ZTE Indonesia, stated that Indonesia is currently entering an important phase of digital transformation, where progress will increasingly depend on strong collaboration between technology providers, infrastructure players, and partners across the digital ecosystem.
“Indonesia is currently entering an important phase of digital transformation, where AI, cloud technologies, and intelligent connectivity will become the key foundations of future digital economic growth. Through ZTE Day Indonesia 2026, we aim to demonstrate how technology innovation can be implemented to support the development of smarter, more efficient, and sustainable digital infrastructure. We believe that cross-industry collaboration will play a crucial role in building a strong digital foundation to support Indonesia’s vision of becoming a leading digital economy,” said Liu Sen.
Beyond showcasing technology innovations, ZTE Day Indonesia 2026 also emphasized the growing importance of ecosystem collaboration in supporting Indonesia’s AI-ready digital landscape.
During the exhibition showcase, ZTE presented a series of its latest innovations, including nubia’s latest AI smartphone, high-performance AI server solutions, optical transport technologies, AI-powered network management systems, and Wi-Fi 7 enterprise connectivity solutions.
Advertisement
ZTE also demonstrated its comprehensive end-to-end digital ecosystem capabilities through solutions covering RAN, microwave, transport network, core network, fixed network, and big video solutions. These innovations reflected the company’s commitment to supporting operators, enterprises, and industries in addressing the evolving demands of the digital era.
As part of ZTE Day Indonesia 2026, the ZTE Open Day Afternoon Session featured keynote presentations from Prof. Viciano Lee of Sertis Indonesia, Sami Muhammad Salman from Whale Cloud Technology Indonesia, Mohan Albert, Director of CTO Group at ZTE, and Chok Shin Lip, Partner Solution Architect Director at Alibaba Cloud Intelligence.
The keynote sessions explored the growing role of AI, cloud technologies, intelligent infrastructure, and ecosystem collaboration in supporting enterprise transformation and accelerating Indonesia’s digital economy development.
The event also hosted a panel discussion titled “Connecting the Ecosystem: Intelligent Connectivity for Enterprise Integration & Value Innovation”, featuring industry leaders including Eric Arianto, Chief Technology & Network Officer of Linknet, Irawan Delfi, Network Development Division Head of Fiberstar, and Sigit Dwi Cahyo, Head of Technology Planning and Product of Tower Bersama Group. The panel explored the importance of intelligent connectivity, fiber infrastructure readiness, and ecosystem integration in supporting enterprise digitalization, service innovation, and the growing demand for seamless digital experiences across industries.
Advertisement
Panel discussion titled “Connecting the Ecosystem: Intelligent Connectivity for Enterprise Integration & Value Innovation”
Another panel discussion titled “Building the Foundation: Digital Infrastructure for Indonesia’s AI Era”, moderated by Vincent Han, featured industry leaders including Abieta Billy from DCI Indonesia, Muljadi Muhali from Fortress Digital Services, and Marlo Budiman from DSST Mas Gemilang. The second panel emphasized the importance of strengthening digital infrastructure readiness, enhancing data center capabilities, and fostering industry collaboration to support the growing adoption of AI technologies and Indonesia’s broader digital transformation agenda.
Panel discussion titled “Building the Foundation: Digital Infrastructure for Indonesia’s AI Era”
Through keynote sessions, panel discussions, interactive product demonstrations, and networking activities, ZTE Day Indonesia 2026 provided customers, partners, and industry stakeholders with deeper insights into AI implementation, intelligent digital infrastructure, and real-world applications of next-generation technologies across industries.
ZTE Day Indonesia 2026 showcased the latest AI, cloud, and intelligent connectivity innovations to support Indonesia’s digital transformation
Image Playground will generate images based on provided photos and other inputs, but the results leave a lot to be desired
Generative AI is not Apple’s strong suit, but Image Playground, the tool responsible for horrific AI avatar generation and Genmoji, should see significant improvements with the OS 27 cycle.
Apple announced Apple Intelligence features that had to be delayed in 2024, and one of them should have been Image Playground. While the other AI tools were of passable usefulness and quality, the image generation tool still leaves a lot to be desired.
According to the “Power On” newsletter from Bloomberg, the Image Playground app should see a “big boost” from Apple’s upgraded Apple Foundation Models. It’s an obvious statement given that Gemini is being distilled into Apple’s models, and one of Gemini’s specialties is image generation.
Advertisement
However, even as Apple’s image generation feature improves, its models will still likely lag behind competing options by some margin. Users that want more oomph behind their image generation in Image Playground will be able to use third-party models of their choosing.
Image generation is bad AI
While the results are often unfortunate, Image Playground is more of a proof-of-concept than a useful tool in its current form. Apple presented it as a way to turn your mom into an AI-slop superhero for what would be a simply terrible birthday message.
While no one on Earth should use Image Playground for that function, it can be somewhat entertaining the same way other gen AI models have been. Using it to see how it might interpret a specific person or prompt can be amusing, but the output isn’t something anyone could or should rely on for anything beyond a giggle in an iMessage group chat.
The only good thing about Image Playground output (when using Apple’s models) is that it’s either on device or in Private Cloud Compute servers that run on renewable energy. While generative slop is loaded with moral quandaries, that’s one area Image Playground doesn’t suffer from.
Advertisement
Genmoji is the true winner in Image Playground. While it isn’t a function I’ve used much, it has enough guardrails that it can make some decent results.
Even if Image Playground is able to one day achieve Pixar or Ghibli-level results with on-device AI, there’s the argument that all gen AI is bad for humanity. One area of image generation Apple will no doubt stay away from is the ability to produce photorealistic output, which can lead to problems like deepfakes.
OS 27 getting new Apple Foundation Models
iOS 27 is expected to make Genmoji more proactive in the system by suggesting premade options in the text suggestion box. Shared Genmoji become available to the people you send them to, so everyone can get some use out of that weed emoji.
Writing Tools are one of Apple’s better AI features
Advertisement
Apple’s weakest offerings in AI are its generative functions, however, they’re fully optional and easily ignored. There is no doubt Apple will continue to pursue better image generation, but if that’s something you want to do, you might want to stick to other apps.
That said, Image Playground could prove much better in OS 27 and earn its place as a useful toolset. The app itself is fine, but Image Playground also exists outside of the app as an extension interface in places like Notes and Freeform.
If Apple can make Image Playground a useful tool that can run on devices with Apple’s arguably somewhat ethical AI models, then I welcome it. Though improvements could mean turning the tool into a plagiarism machine, which will create a whole different set of issues.
People that want to perform these functions are going to seek out tools that offer them. If gen AI must exist to meet consumer expectations, then Apple should have a good offering.
Advertisement
These fake images that lack artistic ingenuity or integrity are getting created no matter what. So, if some of them can be made with ethically sourced AI, local models running on a device’s battery, or on private servers powered by renewable energy, I’d say it is a net positive.
The ability to choose third-party models for Image Playground through a system API will prove interesting. However, once you’ve sent your data to that third-party, all bets are off from an ethical and privacy-oriented standpoint.
At least you can make a plagiarized Ghibli avatar for social media instead of paying a human to make one for you, I guess.
Cisco has released security updates to address a maximum-severity Secure Workload vulnerability that allows attackers to gain Site Admin privileges.
Formerly known as Cisco Tetration, Cisco Secure Workload helps admins reduce their network’s attack surface through zero trust microsegmentation and stop lateral movement to keep business applications safe.
Tracked as CVE-2026-20223, the security flaw was found in Secure Workload’s internal REST APIs, and it enables unauthenticated attackers to access resources with the privileges of the Site Admin role.
“This vulnerability is due to insufficient validation and authentication when accessing REST API endpoints. An attacker could exploit this vulnerability if they are able to send a crafted API request to an affected endpoint,” Cisco explained in a Wednesday advisory.
Advertisement
“A successful exploit could allow the attacker to read sensitive information and make configuration changes across tenant boundaries with the privileges of the Site Admin user.”
Cisco says there are no workarounds for this security flaw, has released software updates to patch it for on-premises customers, and has already addressed it in the cloud-based Cisco Secure Workload SaaS deployment.
Cisco Secure Workload Release
First Fixed Release
3.9 and earlier
Advertisement
Migrate to a fixed release.
3.10
3.10.8.3
4.0
4.0.3.17
The company also added that its Product Security Incident Response Team (PSIRT) has not found evidence that the vulnerability has been exploited in the wild before publishing this week’s advisory.
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) added the CVE-2026-20182 flaw to its Known Exploited Vulnerabilities Catalog on May 14 and ordered federal agencies to secure affected devices within three days, by May 17.
Over the past five years, CISA has flagged 91 Cisco vulnerabilities as actively exploited, six of which have been used by various ransomware gangs.
Automated pentesting tools deliver real value, but they were built to answer one question: can an attacker move through the network? They were not built to test whether your controls block threats, your detection rules fire, or your cloud configs hold.
This guide covers the 6 surfaces you actually need to validate.
CEO eyes margin gains by keeping headcount flat – bold for a company selling HR software to employers
Workday is hoping to boost its revenue
and margins by using AI agents instead of hiring
people, according to its CEO.
After announcing revenue growth, Aneel Bhusri – the company co-founder who was reinstated as CEO in February – said his aspiration is to keep headcount
the same while sustaining growth and increasing margins by harnessing AI.
Advertisement
“I’d love to see us continue the growth
that we had in Q1, but keep headcount as close to flat for the year as possible
because we are getting the benefits of using our own products and other AI
tools. That’s where I’m hopeful and believe that we’re going to have additional
margin expansion as we get those benefits. That’s different than what my view
was coming in three months ago.”
In its Q1 results ended April 30, Workday recorded net profit of $222 million versus $68 million in the prior year, when the bottom line was hit by restructuring expenses. Revenue generated for the three months was $2.54 billion, up 13.5 percent year-on-year.
The results beat market expectations and Workday forecast higher margins for the rest of the year, sending its share price up 10 percent in
after-hours trading.
Bhusri’s aspiration to keep headcount flat
while increasing revenue and margins follows a roller-coaster ride of public statements on employment
plans.
Advertisement
In February 2025, Workday announced
an 8.5 percent cut to its global workforce – 1,750 positions – as it “intended to prioritize its investments and continue advancing Workday’s ongoing focus on durable growth,” an SEC
filing said.
In June 2025, CFO Zane
Rowe told an investment conference that the SaaS biz planned to rehire the same number
of people, although with different roles. “We will be hiring back. We wanted to
make sure everyone understood that this is not us reducing,” he said.
Nonetheless, in September 2025, then CEO
Carl Eschenbach seemingly reversed the plan, telling investors it was “consolidating
and streamlining the organization model” and did not “need more
headcount to drive the business forward.”
Shareholders may be delighted that Workday
can now expand without having to increase the size of its workforce. But for a company that
relies on organizations hiring people to create demand for its HR software, it seems like a strange example to set. ®
You must be logged in to post a comment Login