When most people think about vacuum tubes, they picture big glass bottles glowing inside antique radios or early computers. History often treats tubes as a dead-end technology that was suddenly swept away by the transistor in the 1950s. But the reality is much more interesting. Vacuum tube technology did not simply stop evolving when the transistor appeared. In fact, some of the most sophisticated and technically impressive tube designs emerged after the transistor had already been invented.

During the final decades of mainstream tube development, manufacturers pushed the technology in remarkable directions. Tubes became smaller, faster, quieter, more rugged, and more specialized. Designers experimented with exotic geometries, ceramic construction, metal envelopes, ultra-high-frequency operation, and even hybrid tube-semiconductor systems. Devices such as acorn tubes, lighthouse tubes, compactrons, and nuvistors represented a last gasp of thermionic electronics.

Ironically, many of these innovations arrived just as solid-state electronics were becoming commercially practical. Vacuum tubes were improving rapidly right up until the market abandoned them.

The Pressure to Improve

By the 1930s and 1940s, vacuum tubes dominated electronics. Radios, radar systems, military communications, industrial controls, and the first digital computers all depended on them. But everyone was painfully aware of their problems.

Traditional tubes were fragile, generated heat, consumed significant power, and suffered from limitations at high frequencies. Internal lead lengths created parasitic inductance and capacitance. At radio frequencies and especially microwave frequencies, those unwanted effects made design difficult.

Military requirements during World War II accelerated development dramatically. Radar systems needed tubes capable of operating at VHF, UHF, and microwave frequencies. Vehicle equipment required devices that could withstand punishment. Computers with tubes suffered from frequent failures, took up entire rooms, and needed special cooling equipment, often bigger than the computer. These pressures drove tube designers into an intense period of innovation.

Acorn Tubes: Tiny Tubes for High Frequencies

One of the earliest major departures from conventional tube geometry was the acorn tube. Developed in the 1930s by RCA, the acorn tube got its name from its distinctive shape, which resembled an acorn with wire leads protruding from the base and sides. Unlike ordinary tubes, where the internal elements had relatively long leads, the acorn design minimized lead length to reduce parasitic capacitance and inductance. At high frequencies, this reduction was crucial.

One famous example was the 955 acorn triode. These tubes found use in experimental television receivers, military radios, and laboratory equipment.  Acorn tubes also reflected an important trend in late tube development: engineers were increasingly treating tubes not merely as amplifying devices, but as microwave structures requiring careful electromagnetic design.

The Lighthouse Tube

If acorn tubes were specialized, lighthouse tubes were positively futuristic. Lighthouse tubes abandoned the classic cylindrical glass form almost entirely. Instead, they used stacked disk-like electrodes arranged in a compact coaxial structure. The resulting geometry minimized transit times and parasitic reactances, allowing operation into microwave frequencies.

The tubes vaguely resembled a lighthouse tower. These tubes became essential in radar systems during World War II and the early Cold War period. Some lighthouse designs could operate in the gigahertz range, something impossible for conventional receiving tubes.

Their construction also introduced new manufacturing techniques. Many used ceramic and metal rather than large glass envelopes. This improved heat resistance and mechanical stability while reducing losses at high frequencies.
In many ways, lighthouse tubes represented the transition from classic vacuum tubes and true microwave devices like klystrons and traveling-wave tubes.

Metal Tubes and Ruggedization

Another path of tube evolution focused on durability and compactness. Early tubes used fragile glass envelopes that were easily broken and susceptible to microphonics and vibration. During the 1930s, manufacturers introduced all-metal tube designs. These tubes replaced the glass envelope with a metal shell, improving shielding and mechanical ruggedness.

Metal tubes were particularly attractive for military and automotive applications. Shielding reduced interference, while the smaller physical size allowed more compact equipment layouts.

Hybrid glass-metal constructions also became common. Engineers experimented constantly with new materials and packaging approaches to reduce noise, improve reliability, and extend tube lifespan.

Subminiature Tubes

One of the most impressive developments was the subminiature tube. These tiny devices often looked more like oversized resistors than conventional tubes. Some were less than an inch long and designed to be soldered directly into circuits rather than plugged into sockets.

Subminiature tubes emerged largely from military demands during and after World War II. Proximity fuzes for artillery shells required electronics small enough to survive being fired from a cannon. Traditional tubes would simply shatter under the acceleration.

The resulting ruggedized miniature tubes were shock-resistant and compact enough for portable military electronics. After the war, subminiature tubes appeared in hearing aids, portable radios, test instruments, and early miniaturized computers.

The Nuvistor: The Ultimate Receiving Tube

One of the most interesting late-stage vacuum tube was the RCA Nuvistor. Introduced by RCA in 1959, the nuvistor represented an attempt to create a truly modern vacuum tube for the transistor age.

Unlike classic glass tubes, nuvistors used a compact metal-and-ceramic construction. They were extremely small, highly reliable, vibration-resistant, and capable of excellent high-frequency performance. They also exhibited very low noise characteristics. At first glance, a nuvistor hardly resembles a traditional tube at all. You could easily mistake these for some other component in a metal can.

Technically, nuvistors were excellent devices. They offered superior performance in many RF applications compared to early transistors, particularly in television tuners, instrumentation, and aerospace electronics.

High-end studio microphones also adopted nuvistors because of their low noise and desirable electrical behavior. Some audiophiles still use nuvistor-based equipment today.

But despite their capabilities, nuvistors arrived too late. Semiconductor technology was improving rapidly. Silicon transistors were becoming cheaper, more reliable, and easier to manufacture in large quantities. Integrated circuits loomed on the horizon. The nuvistor may have been the best small receiving tube ever made, but it was competing against a technology whose economics would soon become overwhelming.

Compactrons

As semiconductor electronics advanced, tube manufacturers attempted another strategy: integration. The Compactron, introduced by General Electric in the early 1960s, combined multiple tube functions into a single envelope. A compactron might contain several triodes, pentodes, or diode sections in one package. This reduced component count, simplified wiring, and lowered manufacturing costs for television sets and other consumer electronics. Of course, tubes with multiple electrodes weren’t new. They dated back to at least 1926. However, GE’s aggressive marketing of the brand was an attempt to prevent designers from defecting to the solid-state camp.

In some sense, compactrons were the vacuum tube answer to integrated circuits. Engineers were trying to achieve greater functional density while keeping tube-based designs economically competitive. GE’s Porta-Color, the first portable color television, used 13 tubes, including 10 Compactrons. They usually have 12-pin bases and an evacuation tip at the bottom of the tube rather than at the top.

Compactrons saw widespread use in televisions, stereos, and industrial electronics during the 1960s and early 1970s. But again, semiconductor integration advanced even faster. The battle was becoming impossible to win.

Specialized Tubes Survived

Even after transistors took over consumer electronics, vacuum tubes remained important in specialized fields. Microwave tubes such as klystrons, magnetrons, and traveling-wave tubes continued to dominate high-power RF applications. Radar systems, satellite communications, particle accelerators, and broadcast transmitters all relied on advanced vacuum devices. In some areas, they still do.

A modern microwave transmitter aboard a communications satellite may still use a traveling-wave tube amplifier because tubes can handle very high frequencies and power levels efficiently.

No Instant Win

One misconception about electronics history is that the transistor immediately rendered tubes obsolete after its invention at Bell Labs in 1947. That is not what happened.

Early transistors had many limitations. They were noisy, temperature-sensitive, low-power, and expensive. Tubes often outperformed them in RF circuits, audio applications, and high-power systems well into the 1960s.

For a significant period, designers genuinely did not know which technology would dominate certain markets. Tube designers were still making substantial advances. Nuvistors and Compactrons were not desperate relics; they were serious engineering efforts intended to compete in a changing world.

Ultimately, however, semiconductors possessed overwhelming long-term advantages. Transistors required less power, generated less heat, occupied less space, and could be manufactured using scalable photolithographic processes. Once integrated circuits became practical, the economics shifted decisively. Vacuum tubes could evolve, but they could not shrink into millions of devices on a silicon chip.

The final years of vacuum tube development are often overlooked because history tends to focus on winners. Yet this period produced some of the most elegant and specialized electronic devices ever created. By the late tube era, vacuum tube manufacturing had become quite refined. Engineers could produce tubes with tightly controlled characteristics and surprisingly long operating lives.

Some early transistorized devices still retained subminiature tubes in certain high-frequency or low-noise stages because transistors had not yet surpassed tube performance in every application. This overlap period is often forgotten today. Electronics did not instantly switch from tubes to semiconductors. For years, many systems used both. For many years, a typical ham radio transmitter, for example, would be all solid-state except for the power amplifier finals, which were often a pair of 6146 tubes.

You can, of course, make your own tubes. If you’ve had enough of making your own tubes, maybe try reproducing some of these advanced models.

California Attorney General Rob Bonta announced a $12.75 million settlement agreement with General Motors (GM) over allegations that the company violated the California Consumer Privacy Act (CCPA).

The violations arise from allegations that the car maker illegally collected and sold Californians’ driving and location data to data brokers Verisk Analytics and LexisNexis Risk Solutions, between 2020 and 2024.

The investigation into this activity began in 2024, following media reports about automakers, including GM, sharing driver behavior with insurers.

The data was allegedly collected through GM’s OnStar subsidiary and its “Smart Driver” system and was reportedly intended for driver-scoring products related to insurance.

The American carmaker, which owns the GMC, Cadillac, Chevrolet, and Buick brands, was previously criticized by the U.S. Federal Trade Commission (FTC) for this unlawful data collection, with the government body banning GM from selling drivers’ data for five years.

The Californian authorities said GM failed to properly notify consumers or obtain their consent for this data collection, and retained the data for longer than necessary, even re-purposing it for sale, and making $20 million nation-wide.

“General Motors sold the data of California drivers without their knowledge or consent and despite numerous statements reassuring drivers that it would not do so,” Attorney General Rob Bonta stated.

“This trove of information included precise and personal location data that could identify the everyday habits and movements of Californians.”

The amount of $12.75 million in civil penalties is a record in the state’s history, and the first case of enforcement action focused on data minimization rules.

In addition to the fine, GM is also required to:

• Stop selling driving data to consumer reporting agencies and brokers for five years.
• Delete retained driving data within 180 days unless consumers explicitly consent to retention.
• Ask LexisNexis and Verisk to delete the data they received previously.
• Implement a stronger privacy compliance program and submit regular assessments to regulators.

The officials said California drivers were unlikely to have faced higher insurance premiums as a result of GM’s data sales, thanks to state law prohibiting insurers from using driving data to set rates.

BleepingComputer has contacted GM with a request for a comment on California’s announcement, but we have not received a response by publication time.

AI chained four zero-days into one exploit that bypassed both renderer and OS sandboxes. A wave of new exploits is coming.

At the Autonomous Validation Summit (May 12 & 14), see how autonomous, context-rich validation finds what’s exploitable, proves controls hold, and closes the remediation loop.

Advertisement

Claim Your Spot

It’s depressing and entirely predictable that any move to improve efficiency is always hit by a wave of misplaced anger and full-on stupidity. In this case, I’m talking about Ed Miliband’s plan for tumble dryers.

In the recent government document, Raising standards for household tumble dryers, the bit that’s been getting certain people chomping at the bit is the sentence, “To improve the efficiency of household tumble dryers, the final regulations introduce a new minimum performance standard that phases out inefficient gas-fired, air-vented, and condenser models.”

In certain sectors, this move has been likened to the Soviet era, as it removes choice and forces consumers to buy more expensive heat pump appliances. All of which are bizarre hot takes, and completely miss the point. I’d know, as I review tumble dryers and have the cold facts.

This situation reminds me of the time that the EU introduced a new law that reduced the maximum power of vacuum cleaners from 1600W to 900W. The same kinds of people who are moaning about tumble dryers also moaned about this rule change, saying that homes would be dirtier and powerful cleaners would be banned.

Of course, those people were wrong about vacuum cleaners. Cleaning performance isn’t so much about raw suction power as it is overall efficiency: lower power vacuum cleaners can have as much suction power as those that draw more power, and things like motorised brushbars can improve dust collection without using significant amounts of power. In a shocking twist, people moaning about tumble dryers are also wrong.

Heat pump tumble dryers cost a little more but are a lot cheaper to run

It’s true, a heat pump tumble does cost more than a vented and condenser models. Looking at entry-level models, you’ll probably pay around £40 more upfront for a heat pump dryer than for a vented or condenser dryer.

That’s not a huge amount, but the more important bit of information is how much they cost to run. I review a lot of tumble dryers (all of my best buy tumble dryers are heat pump models), and the cold hard truth is that heat pump models are significantly cheaper to run.

In our guide, condenser vs heat pump tumble dryers, we looked at running costs for a washer-dryer (which uses a condenser for drying) and a heat pump model. The washer dryer used 2.387kWh to dry a load of clothes, and the heat pump used 0.813kWh.

At the current price cap of 24.67p per kWh of electricity, that’s a cost of 58.88p for the condenser dryer and 20.05p for the heat pump. That means that the condenser dryer costs 38.83p per cycle more to run.

In 103 cycles, the heat pump tumble dryer has clawed back that extra £40 it cost. That could be under a year (if you run a couple of cycles per week), or within two years. Everything past that is just more saving.

Several people complaining about the tumble dryer plan have said that they’d focus on reducing energy costs instead. So, say we could get electricity prices back down to around 16p per kWh.

In this scenario, our condenser dryer would cost 37.19p per cycle and the heat pump dryer 13p per cycle. The condenser is still 24.19p per cycle more expensive to run. In this scenario, it would take 165 cycles to get your extra £40 outlay back, but that’s still within two or three years for more people. Given that you’re likely to have a dryer for a good eight to 10 years, you’ll still save money in the long run.

Heat pump dryers do take longer, but they’re better for your clothes

Another thing thrown at heap pump tumble dryers is that they take longer to dry your clothes. That’s true, because of the way that they work.

Vented and condenser tumble dryers work via a heating element that’s passed into the drum at around 70°C, heating your clothing and removing moisture. This moisture is then either condensed into a tank (a condenser dryer) or vented out the back (a vented dryer).

Using hotter air speeds up the drying process, but it requires more electricity and the heat you generate is just pumped out.

Heat pump tumble dryers use a closed system, and use a compressor rather than a heating element. These dryers extract heat from the closed system and recycle it continuously (our guide, What is a heat pump tumble dryer?, explains more), running at around 50°C.

That lower temperature means longer drying times, but less energy use overall. And lower temperatures are better for clothes, as they’re kinder to the fabric and less likely to cause issues such as shrinking.

Aside from the money factor, energy is a precious resource, so why not use it as efficiently as possible?

That’s even the case if you have solar and are generating your own power. With a heat pump, less of your solar goes onto drying your clothes, which means, depending on the size of your array, you’re more likely to be able to generate the full power for the cycle, or your spare capacity can be used elsewhere: topping up a battery, running your TV or just for exporting to the grid, where you’ll get paid for it. Doing more with the resource you have is better, no matter how you cut it.

The only real downside of heat pump tumble dryers is that they are best used in warmer rooms. According to Hoover, its heat pump tumble dryers are designed to work in rooms that are always warmer than 7°C, so they may not work in cold rooms or garages in the winter months.

There are alternatives. Hotpoint ColdGuard Heat Pump tumble dryers are built to work in rooms as cold as 0°C.

Heat pump tumble dryers are better than traditional types, and the government taking steps to ensure that efficient models are the only ones you can buy makes sense in the long run. Here, as in any other area, efficiency is better for everyone and a cost saving is a cost saving, regardless of the price of electricity.

For decades, assessing cholesterol risk has been built around a simple idea: Lower “bad” cholesterol, lower your chance of a heart attack. The test at the center of that approach measures how much low-density lipoprotein, or LDL cholesterol, is circulating in part of the blood. It has shaped everything from clinical guidelines to the widespread use of statins, medications that reduce LDL.

It works. Lowering LDL cholesterol reduces heart attacks, strokes, and early death. But it doesn’t tell the whole story.

The LDL cholesterol test measures the amount of cholesterol inside the low-density lipoprotein particles circulating in the bloodstream. Those LDL particles containing the cholesterol can get trapped in artery walls, forming plaques that can eventually block blood flow. As the test measures the amount of cholesterol being carried, not the number of LDL particles themselves, two people can have the same LDL cholesterol level but very different numbers of particles, and therefore different levels of risk.

That gap has pushed researchers toward a different way of measuring risk. Apolipoprotein B, or apoB, reflects the total number of cholesterol-carrying particles in the blood rather than how much cholesterol they contain. A growing body of research suggests it’s a more accurate way of identifying who is at risk and who’s not.

In March 2026, the American Heart Association and American College of Cardiology recognized this. Their updated cholesterol guidelines acknowledged apoB as a potentially more precise marker, in line with earlier European recommendations. But they stopped short of recommending apoB as the primary method for testing.

“They review the evidence and rank apoB as superior, but the actual rules of the road continue to prioritize LDL,” says Allan Sniderman, a cardiologist at McGill University.

Sniderman was an author on a 2026 JAMA modeling study that analyzed lifetime outcomes for around 250,000 US adults eligible for statin treatment. Comparing LDL cholesterol, non-HDL cholesterol, and apoB, the study found that using apoB to guide treatment decisions would prevent more heart attacks and strokes than current approaches, while remaining cost-effective.

ApoB testing can be done through standard blood tests. So why has it not filtered into routine care? Not even in Europe, where the guidelines have reflected its usefulness for years.

Part of the answer is inertia. For decades, LDL cholesterol has been both a scientific breakthrough and a public health success story. It is simple, widely understood, and directly linked to treatments that work.

“For 50 years, LDL cholesterol was an amazing discovery,” Sniderman says. “It’s not that it isn’t a good marker. It is a good marker.”

Børge Nordestgaard, president of the European Atherosclerosis Society, agrees that LDL cholesterol remains central for a reason. “The evidence is immense; it’s beyond discussion,” he says. “Statins reduce heart attacks, strokes, and early death through LDL cholesterol lowering.”

That success helped shape a powerful narrative: LDL is “bad cholesterol,” and lowering it saves lives. But that simplicity has also limited how risk is understood.

“The result is patients and physicians know little or nothing about apoB,” Sniderman says.

More recent research suggests that the cholesterol picture is more complex, especially in people already taking statins. Previous studies led by Nordestgaard have shown that in treated patients, high levels of apolipoprotein B and non-HDL cholesterol remain associated with increased risk of heart attacks and mortality, while LDL cholesterol does not. ApoB, in particular, emerged as the most accurate marker.

For Kausik Ray, a cardiologist at Imperial College London, the challenge is not choosing one marker over another, but understanding what each one captures, and what it misses.

“We’re not interested in cholesterol for its own sake,” Ray says. “We’re trying to prevent heart attacks and strokes.”

Checkmarx warned over the weekend that a rogue version of its Jenkins Application Security Testing (AST) plugin had been published on the Jenkins Marketplace.

The compromise was claimed by the TeamPCP hacker group, which initiated a spree of supply-chain attacks that included the Shai-Hulud campaigns on npm and the Trivy vulnerability scanner breach, resulting in the delivery of credential-stealing malware.

Jenkins is one of the most widely used Continuous Integration/Continuous Deployment (CI/CD) automation solutions for software building, testing, code scanning, application packaging, and deploying updates to servers.

The Checkmarx AST plugin on the Jenkins Marketplace integrates security scanning into automated pipelines.

“We are aware that a modified version of the Checkmarx Jenkins AST plugin was published to the Jenkins Marketplace. We are in the process of publishing a new version of this plug-in,” Checkmarx alerted in the update.

This is the third incident in a series of supply-chain attacks the application security testing firm has suffered since late March.

According to offensive security engineer Adnand Khan, TeamPCP gained access to Checkmarx’s GitHub repositories and backdoored the Jenkins AST plugin to deliver credential-stealing malware.

A company spokesperson confirmed to BleepingComputer that the threat actor obtained credentials to the repositories from the Trivy supply-chain attack in March.

A message the hackers left in the about section reads: “Checkmarx fails to rotate secrets again. With love – TeamPCP.”

“As a result of that access, the attackers were able to interact with Checkmarx’s GitHub environment and subsequently publish malicious code to certain artifacts,” the company spokesperson stated.

Using credentials stolen in the Trivy attack, the hackers published modified versions of multiple developer tools on GitHub, Docker, and VSCode that included info-stealing code.

The threat actor maintained access for at least a month and then published a malicious version of the company’s KICS analysis tool on Docker, Open VSX, and VSCode, which harvested data from developer environments.

In late April, the company confirmed that the LAPSUS$ threat group leaked data stolen from its private GitHub repository.

On Saturday, May 9, a rogue version (2026.5.09 ) of the Checkmarx Jenkins AST plugin was uploaded to repo.jenkins-ci.org. The update was outside the plugin’s release pipeline and included malicious code.

Apart from not following the official date style scheme, the malicious plugin lacked a git tag and a GitHub release.

Checkmarx advised users to ensure that they are using version 2.0.13-829.vc72453fa_1c16 of the plugin published on December 17, 2025, or an older one.

Although Checkmarx hasn’t shared any details about what the rogue Jenkins plugin does on systems, those who have downloaded the malicious version should assume that their credentials are compromised, rotate all secrets, and investigate for lateral movement or persistence.

Checkmarx says that its GitHub repositories are isolated from its customer production environment, and no customer data is stored in the GitHub repository.

“We have communicated with our customers throughout this process and will continue to provide relevant updates as more information becomes available,” the cybersecurity company said, adding that customers can find recommendations on the Support Portal or in the Security Updates sections.

Checkmarx has published a set of malicious artifacts that defenders can use as indicator of compromise (IoCs) on their envirronments.

AI chained four zero-days into one exploit that bypassed both renderer and OS sandboxes. A wave of new exploits is coming.

At the Autonomous Validation Summit (May 12 & 14), see how autonomous, context-rich validation finds what’s exploitable, proves controls hold, and closes the remediation loop.

Advertisement

Claim Your Spot

I went into this research convinced I already knew the answer.

After more than a decade in People Analytics, the last few years at Meta, I had a working theory about why tech employees leave their jobs within the first year. Two things, I believed, were doing most of the damage: whether someone was getting promoted, and how often they were socialising with their immediate team outside of work. The first felt obvious. The second felt like the kind of human factor the industry consistently underweights.

I was half right.

When I surveyed 205 tech professionals globally and trained a machine learning model to predict early attrition, promotions came out as the single strongest signal in the dataset. But socialisation? It barely registered. And the factors that did matter alongside promotions pointed somewhere I hadn’t fully anticipated. Early attrition in tech isn’t primarily a culture problem. It’s a career momentum problem.

That finding changed how I think about retention. I suspect it might do the same for you.

Tech has always had an attrition problem

The technology industry has one of the highest attrition rates of any sector. Median tenure at many tech companies sits at around one year, regardless of company size. This isn’t a post-pandemic hangover or a hot job market anomaly. It’s been the structural baseline for as long as the industry has existed, and the industry has never really solved it.

The costs are well documented. Replacing an employee can run up to 2.5 times their salary once you factor in recruiting, onboarding, lost productivity and the institutional knowledge that walks out the door with them. Research suggests that a single standard deviation increase in attrition rate correlates with an 8.9% drop in profits. In an era where tech companies are simultaneously pouring billions into AI infrastructure and scrutinising every other line of their cost base, haemorrhaging money on preventable attrition is a harder position to defend than it used to be.

What’s less well understood is why the problem persists despite enormous investment in trying to fix it. Tech companies spend heavily on perks, engagement programmes, culture initiatives and manager training. Some of it works at the margins. None of it has bent the curve in any meaningful way.

Part of the reason, I’d argue, is that most retention efforts are reactive. Someone signals they’re unhappy, or worse, hands in their notice, and the response kicks in. By then it’s usually too late. The question that has always interested me professionally isn’t how to respond to attrition once it’s happening. It’s whether you can see it coming early enough to do something about it.

There was no dataset for this, so I made one

The first problem I ran into was data. There’s no shortage of public datasets on employee attrition, but almost none of them specify industry. The most widely used one is a fictional HR dataset created by IBM data scientists, which has been recycled across dozens of academic studies. It’s clean, it’s accessible, and it tells you nothing specific about the technology sector.

So I built my own. I designed a 24-question survey and distributed it globally to professionals in the tech industry, with one hard requirement: both their current and previous employers had to be technology companies. After removing duplicates and incomplete responses, I had 205 usable records. Not a massive dataset by industry standards, but clean, specific, and purpose-built for the question I was trying to answer.

I defined “early attrition” as leaving a job within the first year. Every respondent was then classified as either an early attrition or not, and that classification became the target the model was trained to predict.

From there, I trained five machine learning algorithms on the data and tested each one across multiple configurations. I utilized an F1 score rather than simple accuracy to measure performance, and the reason matters. A model predicting whether someone left within a year could technically achieve high accuracy just by labelling everyone “stayed longer” since that’s the more common outcome. An F1 score accounts for that imbalance and gives a more honest picture of how well the model is actually working. The best-performing setup combined an algorithm called Extra Trees Classifier with a technique called SMOTE, which addressed the imbalance in the dataset by generating synthetic examples of the minority class. That combination achieved an F1 score of 0.97 out of a possible 1.

The model worked. The more interesting question was what it had learned.

Promotion was the loudest signal in the room

Of all the variables in the dataset, the number of times someone had been promoted in their previous job was the single strongest predictor of whether they left within the first year. The correlation was -0.54, which in plain terms means this: the fewer promotions someone had received, the more likely they were to be an early attrition. Not marginally more likely. Significantly more likely.

This confirmed half of my original hypothesis, and it shouldn’t surprise anyone who has worked in tech. Promotion isn’t just a title change or a pay increase. For most people, especially earlier in their careers, it’s the primary signal that the company sees them and is investing in their future. When that signal doesn’t come, people start looking for it elsewhere.

Nearly half of the respondents in my survey, 49%, had never been promoted in their previous job. That number sat with me. In an industry that prides itself on meritocracy and moving fast, nearly half the sample had never received a single formal recognition of progression. The model was picking up on something that was hiding in plain sight.

Alongside promotions, three other factors emerged as meaningful predictors. Each one is worth unpacking individually because the directionality isn’t always what you’d expect.

Age. Younger workers were significantly more likely to be early attritions. The correlation between age and early departure was -0.49, meaning the older the respondent, the less likely they were to have left within the first year. This makes intuitive sense when you think about it from a career psychology perspective. Earlier career employees carry less sunk cost, face more aggressive recruiting from competitors, and tend to have higher expectations of rapid progression. When those expectations aren’t met quickly, they move. For HR leaders, this means early-career and new graduate hires deserve disproportionate attention in the first twelve months. Visible career pathing and early promotion signals aren’t a nice-to-have for this cohort. They’re retention infrastructure.

Internal role changes. This one cuts against a common assumption. Employees who had experienced more role changes within their previous company were actually less likely to have been early attritions, with a correlation of -0.49. The instinct is often to treat internal mobility as a sign of restlessness. The data suggests the opposite. Movement inside a company appears to be a marker of engagement and investment, not instability. People who get moved around, who change teams or functions, are people who have been given reasons to stay invested. Rotational programmes and internal transfers aren’t just good for skill development. They’re retention tools.

Manager changes. The most counterintuitive finding in the dataset. Employees who had experienced more manager changes in their previous company were less likely to have left within the first year, with a correlation of -0.44. The assumption most people make is that manager instability drives attrition, and there is plenty of research supporting that at a general level. But within this dataset, the relationship ran the other way.

One thing worth being transparent about across both of these findings: someone who left within the first year simply had less time to accumulate role changes or manager changes than someone who stayed longer. That tenure effect is real and worth acknowledging. But the directional signal still holds. Employees who had weathered multiple manager changes or moved across teams had, by definition, found reasons to stay through organisational disruption. They had built enough roots that a change in reporting line or a shift in scope wasn’t enough to push them out. The dependency on any single manager or single role appears to be highest in the early months, before an employee has built broader organisational depth.

Taken together, these four factors point toward a consistent underlying pattern. Early attrition in tech tends to cluster around employees who are younger, less promoted, less mobile internally, and less embedded in the organisation. They haven’t been stagnant necessarily, but they haven’t been invested in either. The model wasn’t identifying people who were inherently likely to leave. It was identifying people who hadn’t yet been given enough reasons to stay.

The socialisation hypothesis didn’t survive contact with the data

I want to be honest about the part of my hypothesis that was wrong, because I think it’s actually the more instructive finding.

My original assumption was that how frequently employees socialised with their immediate teammates outside of work would be a meaningful predictor of early attrition. The logic felt sound. A sense of belonging, of actually liking the people you work with enough to spend time with them voluntarily, seemed like exactly the kind of human glue that keeps people in their seats during the first year when everything else is still uncertain.

The data didn’t support it. Socialisation frequency came out as one of the weakest signals in the entire model, with a near-zero correlation to early attrition after balancing the dataset.

I’ve thought about why that might be. One possibility is that socialisation outside work is a symptom of a good job rather than a cause of staying in one. People socialise with their teams when things are going well, when they feel settled, when they’re not spending their evenings on LinkedIn. It may be more of a trailing indicator than a leading one. Another possibility is that within the specific context of the first year, career momentum simply carries more weight than social connection. You can like your team and still leave if you’re not being promoted, not being moved, not being invested in.

What this tells me, practically, is that companies leaning heavily on culture and social programming as a retention strategy may be solving for the wrong thing, at least in the early tenure window. Those investments aren’t wasted. But if promotion cadence and internal mobility are broken underneath, no amount of team offsites is going to close the gap.

The signal is already in your data

Here is what I find most striking about these findings. None of the factors the model identified as predictive of early attrition are hidden. They’re not buried in sentiment data or detectable only through expensive listening programmes. Promotion history, age, internal mobility, manager changes. Every one of those data points exists in your HRIS right now. Most companies are sitting on the signal and not reading it.

The pattern the model learned to recognise looks something like this. An employee who is earlier in their career, has been in role for several months without a promotion conversation on the horizon, has never moved teams or changed scope internally, and whose entire organisational identity is still tied to a single manager they may or may not have a strong relationship with. That person is not necessarily unhappy yet. They may not have even consciously decided to leave. But the conditions for early attrition are already in place.

The traditional response to that situation, if it gets noticed at all, is reactive. A skip-level conversation after someone flags dissatisfaction. A retention offer after a competing offer has already landed. A manager coaching conversation after the engagement survey comes back low. By that point the decision is usually already made, or close to it.

What the data suggests is that the intervention window is much earlier than most organisations treat it. The first six months of someone’s tenure is when the pattern is being set. Are they getting feedback that signals a future at this company? Are they being considered for stretch opportunities or cross-functional projects? Is someone actively managing their career trajectory, or are they simply being left to onboard and get on with it?

This doesn’t require a machine learning model to act on. It requires People Analytics teams and HR business partners to start treating early tenure as a risk period that deserves structured attention, not just a probationary formality. Simple cohort analysis on your existing workforce data can surface who is sitting in the high-risk pattern right now. Who is under 30, has been in their role for more than eight months, has never changed teams, and has not had a promotion discussion documented? That list exists in your data today.

The AI investment angle matters here too. At a moment when technology companies are making significant bets on artificial intelligence and scrutinising headcount and operational costs more carefully than they have in years, the economics of preventable attrition look different than they did in a looser environment. Losing an early-career employee within the first year and absorbing the cost of replacing them, which can reach up to 2.5 times their salary, is not just a talent problem. It is a financial inefficiency that sits alongside every other cost a leadership team is being asked to justify.

Retention, viewed through that lens, stops being a soft HR metric and starts looking like an operational priority.

The harder question isn’t who’s leaving. It’s who you’ve given a reason to stay

Predicting attrition is the easier half of the problem. I want to be clear about that. A model can learn to recognise the pattern of someone who hasn’t been invested in. What it can’t do is tell you why your organisation keeps producing that pattern, or what it would actually take to change it.

That’s the question I’d leave with every HR and People Analytics leader reading this. Not “how do we build a model like this” but “what would we find if we ran this kind of analysis on our own workforce today?” Because the data is there. The pattern is legible. The gap is almost always in whether anyone is looking for it with enough time to act.

Tech companies are currently navigating one of the more complicated cost environments the industry has seen in a while. AI infrastructure spending is accelerating at a pace that is putting real pressure on every other budget line. Headcount decisions are being made with more scrutiny. The tolerance for inefficiency, financial or operational, is lower than it has been in years. In that context, the cost of losing an employee within the first year and absorbing the full weight of replacing them sits in uncomfortable tension with the AI investment conversation happening in the same leadership meeting.

You cannot cut your way to efficiency while quietly haemorrhaging talent at the bottom of the tenure curve. The two conversations need to be in the same room.

The research I did was a starting point, not a solution. A survey of 205 professionals, a machine learning model, a set of findings that confirmed some assumptions and challenged others. What it pointed toward, more than anything, is that early attrition in tech is not mysterious. It follows a pattern. That pattern is detectable. And in most organisations, the data needed to detect it already exists.

The question is whether anyone is looking.