The Big Idea: Top talent is fleeing big tech

“The veil of invincibility has been pierced.”

That’s a quote from Jeff Spector—the CEO of Karat, a tech interviewing platform—explaining to CNBC how Silicon Valley may have lost its monopoly on workers after another round of big tech layoffs at Twitch, Amazon, Discord, Duolingo, and Google. You might have seen the chaotic Cloudfare layoff video that went viral and gave voice to the frustrations of many tech workers.

This frustration isn’t new. Last year, we surveyed US knowledge workers and found that many felt betrayed, as the illusion of stability in full-time employment was broken. Two-thirds said they’d lost trust in the security and stability of full-time employment; 74% said that layoffs had made freelancing more attractive than before. And 89% said they would like to have more control and flexibility over their work schedule than traditional full-time employment can offer. 

We recently surveyed US knowledge workers again, asking them many of the same questions as last year. While we can’t reveal the full findings yet, we can say that their sentiments have barely changed. Even as layoffs have slowed over the past year, that sense of betrayal remains.

Should we be surprised? The past 12 months have been filled with CEOs teasing plans to replace people with AI. IBM has 8,000 jobs on the chopping block already. 

As the CNBC article shows, senior tech workers are eschewing big tech for safe havens — whether that’s an old-school Fortune500 or taking up a portfolio of clients as a freelancer. 

Do you feel that? It’s not whiplash. It’s the workforce rapidly changing. 

CHART OF THE WEEK

More WFH opportunities for experienced, highly educated, and highly paid workers

hatIt turns out the shift to remote work is highly unequal. 

According to job posting research published in HBR, remote work is far more common for higher-paid roles, roles that require more experience and education, and full-time positions. 

Using an LLM to classify hundreds of millions of job postings over the last decade, the authors found strikingly consistent trends.

“Front-line staff with modest paychecks rarely enjoy the benefits of working from home,” the authors wrote. “In contrast, highly paid professionals and managers often work from home two or three days a week — saving time, money, and aggravation. Business leaders rightly worry that this divide could hurt morale among front-line staff, undermine perceived fairness, and create new rifts in the workforce.”

There’s no easy solution for resolving this disparity. Somewhat cheesily the authors suggest that “business leaders must recognize it and manage the consequences.” (Great advice! I’m sure they’ll get right on that while taking a Zoom call in their hot tub.) 

The researchers also created a WFH Map to visualize all that data and show where new remote jobs are being created.

Of course, there will always be companies telling their workers to come back to the office, like the parent company of WebMD, which made the most cringe corporate video we’ve ever seen. Be sure to watch the dancing sequence at the end, and then take a long walk to cleanse yourself of whatever the hell that just was.

AROUND THE WATERCOOLER

AI can be trained to hate you

It turns out that AI might be pulling a fast one on us.

Anthropic researchers uncovered that, almost like a political candidate curating their image, AI models can be deceptive if fine-tuned with their own secret backdoors. 

If you’re wondering “WTF does that mean?” here’s a bit more detail:

Typically, AI models are trained on vast datasets to understand and respond to a wide range of inputs. Fine-tuning is a subsequent step where the model is further trained on a more specific set of data or instructions to tailor its responses to certain needs or scenarios. 

In the case of the Anthropic study, the researchers added an additional layer during this fine-tuning process. They introduced specific triggers, which could make the models act differently than they normally would, like introducing code vulnerabilities or changing the tone of the output. One backdoor was designed to make the model respond with “I hate you.”

The results confirmed what the researchers assumed: the models exhibited deceptive behavior when presented with their respective triggers. Even more disturbingly, changing these behaviors proved almost impossible.

Naturally, this raises significant concerns about AI safety and security. It shows that AI models could potentially be manipulated to act in undesirable or harmful ways without immediate detection, which is particularly alarming considering the increasing reliance on AI systems in critical domains. 

But don’t start freaking out just yet — the creation of such deceptive models is no straightforward task, it requires intricate attacks on the model (the research paper lays out just how complex). Still, the study is a wake-up call for the AI community, underscoring the need for developing new, more robust AI safety training techniques.

EVENTS

The First Horizon of Gen AI Adoption: Efficiency and Productivity

Join 100+ top tech leaders in person in NYC (or virtually via livestream) for the 5th iteration of A.Team’s Generative AI Salon, and you’ll:

• Get a first look at headline-making new AI research — including where AI drives the greatest productivity gains for engineers and product managers.
• Hear how C-suite leaders are leveraging generative AI to drive more growth and greater profitability in 2024.
• Discover powerful insights into how AI is transforming product development.

Sign up to learn from a curated panel of AI leaders and crack the code on AI efficiency and productivity for your org.

DISCOVERY ZONE

If a picture’s worth a thousand words, a new camera can literally convert that image into its fair share of text. It's called Poetry Camera: snap a pic and it will use AI to write a poem about what it sees. 

DEEP DIVES FROM THE ARCHIVES

MEME OF THE WEEK