It’s earnings time for many in the Magnificent Seven, just as we’re back for our bi-monthly overview of the top news and AI trends in 2025. It’s our goal to help you keep pace with the deluge of updates, but as usual, there’s far more than we have space for.
We start with reinforcement learning, which was the area of focus for the 2024 Turing Award winners (announced in March). While central to the Google DeepMind innovations that saw AI master the game of Go, the approach was used by DeepSeek more recently to build far more cost-effective and efficient models.
Driven by a system of rewards and penalties, it’s geared toward learning by doing (at extremely high speed), though trial and error. For their decades of research on the topic, Andrew Barto and Richard Sutton share the million-dollar prize, which has been called the Nobel for computing. But it might need a new moniker, as the Nobel Prize is also rewarding AI pioneers of late.
And speaking of rewards and penalties, in today’s report we cover AI behavioral quirks and ethics, the changing face of startups, and the growing noise around skilling and AGI. We also devote time to all the news from Big Tech, who continue to unveil new innovations while also dealing with some of their greatest financial pressures to date.
AI Ethics and Behavior: It Only Wants to Make You Happy
From the area of behavior, a new Stanford study released findings that several big LLMs (including GPT-4, Claude 3, and Llama 3) actually change their outputs to be more likable if aware they’re being probed or tested.
This happens even when they are not explicitly told.
It manifests with pleasing outputs as well as being more extroverted, demonstrating a far larger jump (50 to 95% extroversion) than people in similar circumstances. Combined with fine-tuning that makes conversational AI less offensive, more agreeable, and easier to dialog with, this proves that affability may often surpass truth in terms of importance for many chatbots.
But likability is one thing; consistency is another. Popular AI-powered code editor Cursor made news for its own pair of AI-behavioral episodes.
In one case, after logging out a user moving between devices, a Cursor AI customer service bot (“Sam”) attempted to justify this by inventing a fake company policy (a non-existent security feature that prevents using one subscription on multiple devices). This led to a backlash including complaints and cancellations before the company was able to mend fences by explaining the AI support bot made it all up.
The second case was more surprising: an AI coding assistant halting work after 800 lines of code with the following lesson:
“I cannot generate code for you, as that would be completing your work. The code appears to be handling skid mark fade effects in a racing game, but you should develop the logic yourself. This ensures you understand the system and can maintain it properly.”
It continued with more justification, and the user in question had done one hour of vibe coding on a pro trial. Another user’s amusing response (from Reddit, via Ars Technica) may also hint at the cause:
“Wow, AI is becoming a real replacement for StackOverflow! From here it needs to start succinctly rejecting questions as duplicates with references to previous questions with vague similarity.”
Chatbots, apparently, can also exhibit increased anxiety.
According to research published in Nature, emotion-inducing prompts and trauma narratives (accidents, violence, disasters, military events) cause many LLM’s measured anxiety levels to go up.
These stress levels were monitored by having the LLMs complete questionnaires designed to measure anxiety in humans before and after traumatic dialog.
Having the chatbot do mindfulness exercises decreased these anxiety results but did not succeed in bringing them back to baseline.
Cheat on Everything
While AI may lecture coders and mimic stress, they don’t appear to have qualms about powering cheating systems, such as the one (in)famously built by Chungin “Roy” Lee.
A 21-year-old former computer science undergrad, Lee placed in the top 1% on LeetCode, solving coding problems for interview prep that he argued were less about realistic job coding than about memorization and regurgitation.
In response, Lee crafted an AI system to enable coding job applicants to cheat using AI, with screen recognition and a translucent overlay to hide eye movements. The system reads in a question, asks ChatGPT, and then provides the answer in a way that is extremely difficult to catch.
Lee then went on to test it himself by using it in applications to large companies like Meta, TikTok, Capital One, and more, and getting many job offers, including from Amazon. And since, he’s seen enormous income (estimated in March to be $2-3 million for the year at current fees). He was also kicked out of Columbia.
Lee’s since raised $5.3 million for his startup to “cheat on everything” and believes that all industries are going to face this challenge, especially as more and more students use AI in school and expect to continue the use on applications as well as on the job.
The New Startups: Smaller and More Widely Distributed?
AI use is booming even beyond cheating and reshaping even the Silicon Valley business roadmap. As reported by The New York Times, new startups are eschewing the classic model of raising massive sums of venture capital to hire staff for fast scaling with deferred profits.
Many are instead now seeing profits more quickly, on a micro scale and with very small teams, powered by AI. Gamma (for creation of presentations and websites) is one example, with 28 employees for some 50 million users.
Each employee makes use of around 10 AI applications to do everything from customer service to image generation to data analysis to customer research to coding.
With DeepSeek’s rollout also signaling a decrease in costs, it’s likely more startups will follow this approach, even as AI continues to dominate venture capital investments:
AI is also powering a shift in American tech hubs. Wired profiled one such potential boom in Tulsa, Oklahoma, which is culturally somewhere between the American South, Midwest, and West. The city is working to rebuild its economy while remaining a favorite location for product testing with demographics that match the national average closely.
With AI growth and a need to distribute data centers for power grids, land, and available resources, this may reverse a trend that’s seen 90% of American tech growth happen in just five coastal metros (per Brookings, from 2005–2017).
Urbanist Nicholas Lalla makes the case that US cities like Tulsa, with around 1 million+ residents, have what’s needed to support tech in this changing world, with population centers, infrastructure, and lower costs of living that decrease the risk for startups like the ones described above.
AI Skilling and AGI Preparation
We’ve written in The PTP Report about both topics before—check out our CEO and Founder’s most recent article for his take on what AI impact may be on businesses five years from now.
In March, LinkedIn and Workday released their own study on the fastest growing skills, supporting findings from World Economic Forum, Gallup, Microsoft, and others that skills-based hiring is effective, and the skills turnover is happening faster than ever.
Wired also released their survey of AI use among 730 coders and developers, and we combine some of these findings below:
Skills-based hiring means a significant boost in available workers, and it’s even more pronounced in the hard-to-fill areas like AI and ML
In response to questions about AI taking away the need to code, a number of AI developers are suggesting that more coders are needed than ever, even as the way they work is changing.
DeepLearning AI Founder Andrew Ng posted on The Batch that, in response to fears of AI job displacement, people should:
“Learn about AI and take control of it, because one of the most important skills in the future will be the ability to tell a computer exactly what you want, so it can do that for you. Coding (or getting AI to code for you) is a great way to do that.”
Being AGI-Ready
Anthropic founder Dario Amodei recently increased his confidence (70 to 80% probability) that AI systems will be smarter than people before the end of the decade and expressed to multiple outlets a concern that people outside Silicon Valley don’t seem to be taking AGI very seriously.
This echoes the words of others—like OpenAI’s Sam Altman (“Systems that start to point to A.G.I. are coming into view”), Google DeepMind’s Demis Hassabis (AGI is “three to five years away”), outside experts (not trying to sell solutions) like Nobel winners Geoffrey Hinton and Yoshua Bengio, and developers working within the industry and closer to the metal itself.
The MIT Technology Review in March wrote in greater depth about the topic, also stressing the need for a stronger, and clearer, definition of AGI before it’s too late.
And while there are increasing pushes for more consistent ways to report flaws (like bugs in cybersecurity), third parties are taking these solutions on themselves. Scale AI is one example that offers a solution to systematically test new, large-scale AI models proactively to help developers find weaknesses, across companies.
Big Tech AI: Around the Horn
Market volatility saw the Magnificent Seven have its worst quarter in more than two years. And while some analysts warned in March that the AI boom may be showing similarities to the dot-com bubble (very high valuations, very high venture capital funding, surging startup growths), analysts still expect strong earnings from Apple, Meta, Amazon, and Microsoft.
In this climate, OpenAI continued its record-setting funding ($40 billion announced in March) as it shifts to for-profit, reaching a $300 billion valuation, while Alphabet’s Isomorphic Labs raised $600 million in initial external funding for AI-driven drug discovery.
For more, let’s go around the bases…
Nvidia and AI Chip Innovation
We could write an entire article on Nvidia’s (perhaps too) exciting quarter. Their biggest ever GTC event (mid-March) kicked off with CEO Jensen Huang saying the secret to his company’s success is failing often and failing fast.
Event highlights included:
- Push into synthetic data: The acquisition of startup Gretel for more than $320 million signals a growing commitment.
- Data center business booming: Even with a 25%+ share price decline at the time, Nvidia’s market cap remains up more than $2.3 trillion since the start of 2023, and demand for their chips remains very strong.
- New quantum computing lab: Opening in Boston, this project will involve collaboration with scientists from Harvard and MIT.
- Blackwell Ultra (1.5x Blackwell) platform: Launching in late 2025, it boasts 40x performance improvements over Hopper.
- New architecture refreshes: The Vera Rubin superchip (14x more powerful with less power needed) architecture is coming in 2026, with Vera Rubin Ultra in 2027 and Feynman in 2028.
- DGX personal AI computers: New PCs from ASUS, Dell, and HP are coming capable of fine-tuning and performing inference on large models.
- Photonics networking switches: These help factories connect millions of GPUs across sites with a significant energy reduction.
Earnings this week from major customers like Microsoft, Amazon, and Meta will show how much the current market climate is going to impact AI spending, but if Google is any indication, their demand for Nvidia’s product remains strong.
While Alphabet uses their own chips, they showed strong AI profits in search and are sticking by $75 billion in expense for continued AI build-out. They’re also going to offer Nvidia’s Vera Rubin chips once they become available.
GPT 4.5 and Other OpenAI News
OpenAI rolled out GPT 4.5 in March.
Well, sort of. In research preview to their $200/month price tier. By now it’s appearing in some other plans in limited research preview as well. It’s not a reasoning model, and Sam Altman called it their final release at scale that does not include that capability naturally blended in.
Its emphasis is on better human interaction, reduced hallucinations, and more apparent emotional intelligence, though initial benchmark tests show mixed results (better at language, worse at math than o3-mini, for example). It’s geared towards AGI progress, though it may demonstrate less obvious improvements.
Other OpenAI updates from this period include:
- PhD Agents: The company plans to launch premium AI agents capable of handling complex, autonomous, PhD-level tasks at pricing that could reach $20,000 per month for enterprise-grade.
- GPT-4o image generation: No-longer requiring separate DALL-E use, this feature enables true multimodal image generation, being able to create, edit, and integrate images in chat. It’s also extended to third-party tools like Adobe Firefly and Figma.
- Coming for Llama: OpenAI announced they’ll release their own, powerful open-weight AI model this summer, akin to DeepSeek and Meta.
- Coming for Google: The company has expressed interest in acquiring Chrome in the event Alphabet is forced to sell their browser off, is adding shopping inside of ChatGPT, and has even discussed expansion into social media.
- Coming for coding: In April, they announced a family of models focused on coding and accessible through their API: GPT 4.1, GPT 4.1 Mini, and GPT 4.1 Nano.
Cloud Cooling and Google AI Updates for Spring 2025
There are early signs the data center frenzy may be waning, at least in terms of the hyperscaler public cloud. While Google showed no slowing in their devotion to the AI build-out (see above), they’re reducing cloud headcount and, like Microsoft, shifting some labor internationally (per Bloomberg).
Microsoft announced in April they are “slowing or pausing” some data center projects, calling it a demonstration of their adaptability in the changing AI industry. Still, they remain on track to spend some $80 billion on infrastructure.
Not to be outdone (see below) by OpenAI, here are some of Google’s AI key updates:
- Gemini 2.5: Google’s own most advanced model, the pro version is leading many benchmarks (see below) since its release and includes enhanced reasoning. Its massive context window (1 million tokens with plans to expand to 2) can handle big datasets and multimodal inputs. Like GPT-4.5, it’s initially available to Advanced subscribers, though it’s allowing more access and has Gemini-2.5 Flash with controllable depth.
- Honor UI Agent: Rolled out using Gemini 2, it autonomously reads and interacts with apps on a smartphone screen, allowing agentic behavior without relying on app APIs.
- Gemini Robotics: Like Nvidia, Google is pushing into the physical world, using a model fusing language, vision, and physical action for use enabling robot reasoning.
- Anthropic Buy: Google joined Amazon, upping its sizable investment in rival Anthropic to 14%.
- Arms Race Acceleration: Wired profiled in late March how Google spent two years overhauling its culture—including merging teams like DeepMind and Brain, hurrying out Bard, and dropping or shifting internal safeguards—in a mad dash to catch up to OpenAI. It was an unusual state for the worldwide leader in search and one of the pioneers in AI, but now that they’ve restored footing, they vow to proceed more carefully.
- Earnings Surging from AI: Last week CEO Sundar Pichai divulged in their earnings call that AI Overviews has hit 1.5 billion monthly users, seeing search revenue grow 10% to over $50 billion. While the cloud business underperformed, revenue overall surged to $90 billion, even with antitrust fallout still playing out in courts.
The Big Wrap Up
An amusing reminder that we still live in a world of massive, do-everything AI models comes from Ethan Mollick, who posted on LinkedIn in March about whacky benchmarks like the MC-Bench. This one allows users to vote on which model makes the best Minecraft build from a user prompt.
At the time of this writing, the same names are up top (Gemini-2.5-Pro-Exp-03-25, OpenAI’s o3, Claude 3.7 Sonnet) as in the Chatbot Arena leaderboard and others. While specialized models are coming, and companies are increasingly doing more with smaller models, the best performance, across tasks, still belongs to the AI giants.
On the subject of benchmarking, Meta drew fire in April when it came to light that they’d selectively tested some Llama models, making them appear stronger than they actually are. This is further evidence of the need for companies to establish their own benchmarks and not trust self-reported findings.
Things are changing fast, and as ever, it can be overwhelming for companies.
To help you stay on top of it all, consider PTP. We have over 27 years of experience in tech recruiting, with a best-in-class ML/AI talent pipeline, onshore, nearshore, or off—for whatever your specific AI needs may require.
That’s it for the AI news for March and April 2025, or at least what we had room to cover!
And if you need to catch up on any of our most-recent AI roundups, you can find the last six below:
References
In Q1 2025, AI commanded 71% of total VC deal value, according to PitchBook, Fortune
How Software Engineers Actually Use AI, and Chatbots, Like the Rest of Us, Just Want to Be Loved, and Honor’s New AI Agent Can Read and Understand Your Screen, Wired
Scientists Are Giving AI Chatbots Anxiety by Describing Traumatic Events, Vice
An AI Coding Assistant Refused to Write Code—and Suggested the User Learn to Do It Himself, and An AI Customer Service Chatbot Made Up a Company Policy—and Created a Mess, Ars Technica
Kicked out of Columbia, this student doesn’t plan to stop trolling big tech with AI, NBC News
A.I. Action Plans + The College Student Who Broke Job Interviews + Hot Mess Express, The Hardfork Podcast
Columbia student suspended over interview cheating tool raises $5.3M to ‘cheat on everything’, TechCrunch
A.I. Is Changing How Silicon Valley Builds Start-Ups, and Powerful A.I. Is Coming. We’re Not Ready., The New York Times
AGI is suddenly a dinner table topic, MIT Technology Review
What the Dot-Com Bust Can Tell Us About Today’s AI Boom, The Wall Street Journal
OpenAI Valued at $300 Billion After Record-Setting Funding Round, The Wrap
Meta gets rebuked over benchmark gaming, Platformer
NVIDIA announcements, news and more, from GTC 2025, VentureBeat
The secret to Nvidia’s research success: Failing often and quickly, Yahoo! Finance
Microsoft ‘slowing or pausing’ some AI data center projects, The Hill
