Welcome to Issue #59

Coming Up:

Mid-2025 AI Predictions Check-In: What Is Actually Happening

Why AI Detectors Fail Students and Teachers

Meta’s V-JEPA2: A New Approach to World Modeling

Plus, we discuss an AI-predicting centaur, how humans might respond to personal proxies in the workplace, AI and art, and all the news we found interesting this week.

It’s Sunday morning!

Robots will soon be showing up on streets around the world and this is the moment in time where we go from that being extremely odd to very normal.

Future grandparents will tell their grandchildren about how there was a time when humanoid robots weren’t everywhere.

But until then, please enjoy this week’s issue.

The DAS Crew - Andy, Beth, Brian, Eran, Jyunmi, and Karl

Mid-2025 AI Predictions Check-In: What Is Actually Happening

At the start of 2025, the predictions came fast. The year of agents. Advances in edge AI. Healthcare breakthroughs. A rise in pushback against AI adoption. Now, past the halfway mark, it is clear which predictions are on track and where expectations need to adjust.

Agents are defining 2025. The promise was that AI agents would shift from demos to real workflows, and that is playing out. From Lindy and Crew AI to the growing use of GenSpark and internal enterprise agents, agentic workflows are handling research, reporting, and operational tasks. We are seeing both single-agent orchestrators with deep context windows and mixtures of smaller task-based agents. Behind the scenes, this is architectural experimentation. For users, it simply means better agents that can take a goal and run with it.

Edge AI is scaling. Adoption is up, with a 20% growth in 2025. Edge AI is moving into manufacturing, energy, and field deployments, bringing models to single-board computers and low-power devices in the wild. Advances in light-based data and chips from companies like Halo are enabling AI to run locally, creating resilient, battery-powered systems for disaster response and field monitoring.

Healthcare sees quiet transformation. While no single “AI cured cancer” headline has landed, we see a steady stream of stories where AI surfaces overlooked patterns, suggests new tests, and helps patients find answers after years of unclear symptoms. People are beginning to expect AI in the loop for healthcare, using tools like ChatGPT for triage and planning conversations with their doctors.

Superintelligence talk rises, AGI remains a moving target. As models show capabilities that would have been called AGI just two years ago, the definitions are shifting. OpenAI’s moving goalposts are a feature, not a bug, of the competitive landscape. Research labs are advancing models that improve themselves, like DeepMind’s Alpha Evolve, pointing toward self-improving systems that will fuel the next debates over AGI and superintelligence.

Pushback is real. Resistance is growing, from creatives who want non-AI options in their tools, to wider cultural skepticism about AI’s role in the economy. Layoffs in sales and marketing, where AI automates high-effort, low-leverage tasks, feed concerns about job loss. The conversations around AI ethics, deployment, and governance are intensifying.

Why AI Detectors Fail Students and Teachers

AI detectors have become a go-to tool for schools trying to police plagiarism, but they are creating more problems than they solve. Tools like GPTZero promise to catch AI-written essays, but they often misfire, tagging original student work as AI-generated. Non-native English speakers and students with different writing styles get flagged unfairly, creating stress and blocking opportunities.

The bigger issue is that the old rules do not fit the new reality. Tools like calculators, spell check, and the internet all went through periods of fear before being integrated into learning design. AI tools are the next evolution, but schools have not caught up. Policies still treat AI as cheating, without clear guidance on when using AI is acceptable, and what skills students should build alongside these tools.

AI detectors are unreliable at the sentence and paragraph level, only gaining some predictive value when analyzing large blocks of text. Even then, false positives remain a risk. Meanwhile, students continue using AI as a learning aid, and professors use it for their own work. The gap between institutional policy and practical reality is growing.

If education is about learning, not just enforcing rules, then AI use needs to be addressed openly. Students should know when using AI is allowed, and what learning objectives they need to demonstrate with or without these tools. Oral assessments, iterative discussions, and human-to-human evaluation may need to return as core methods for student grading and evaluation, insuring students understand, rather than simply checking off their attendance and submissions requirements boxes.

WHY IT MATTERS

AI Literacy Becomes Essential: Schools need to teach both students and teachers what AI can and cannot do, and where its use is appropriate.

Old Tools Cannot Police New Realities: AI detectors misidentify original work, leading to unfair penalties and lost opportunities. Beyond reasonable doubt is a standard not yet practiced when educators are using AI detectors.

Focus Shifts to Outcomes: If the goal is to ensure understanding, education systems need to design processes that confirm learning, not just catch potential misrepresentation of authorship.

Students Already Use AI: Policies ignoring AI do not stop its use; they only create confusion and inconsistency in enforcement.

Opportunity to Rethink Learning: AI can be used to build curiosity and deeper understanding if integrated with clear guidelines and human engagement.

Meta’s V-JEPA2: A New Approach to World Modeling

Large language models have limits. They process text well but struggle to understand how the physical world actually works. Meta’s new V-JEPA2, or Video Joint Embedding Predictive Architecture 2, takes a different path. It is a world model that learns physics and motion by watching curated YouTube videos, not by reading words.

V-JEPA2 trains on millions of hours of video to learn how objects move, fall, collide, and interact. It predicts what will happen next, not at the pixel level, but by understanding relationships and changes over time. This approach to training makes it 30 times faster than Nvidia’s similar models, while using less data to learn complex actions.

Instead of needing detailed programming or thousands of trial runs, robots using V-JEPA2 can watch and learn. They can see a coffee cup on the edge of a table and predict it will fall. They can see a fridge door open and know what likely happens next. This predictive ability is a core step toward making robots useful in real-world, unpredictable environments.

Meta’s open-sourcing of V-JEPA2 also matters. Researchers, small labs, and even hobbyists can test it on home robotics, testing predictive learning without requiring the compute power of a major lab. This makes it possible to explore robotics use cases from disaster response to elder care and warehouse automation.

Leaders like Yann LeCun believe models like V-JEPA2 will be key to moving beyond the limits of language models toward systems that can reason, plan, and act in the physical world.

WHY IT MATTERS

Physics Beats Text: V-JEPA2 learns from watching videos of the real world, building models of motion and cause-and-effect that LLMs cannot match.

Robots Gain Useful Prediction: Predicting what happens next makes robots safer and more capable in environments where people and objects can move suddenly or unpredictably, and fast reaction is needed.

Training Speed and Efficiency Improve: By focusing on relational understanding instead of pixels, V-JEPA2 trains faster while using less compute.

Open Access Enables Innovation: Researchers and builders can experiment with world modeling without massive hardware budgets.

Pathway to True Embodied AI: Moving from text-training to world-understanding is key for robotics, AR systems, and assistants that work with humans in real life.

Just Jokes

Did you know?

Scientists at the Helmholtz Institute for Human‑Centered AI have developed a “psychology” model named Centaur that can predict human behavior with 64 percent accuracy, even in brand‑new situations. Centaur was trained on a massive dataset: 10 million decisions from 60,000 people across 160 psychology experiments. It can forecast actions and reaction times, not just make guesses.

What sets Centaur apart is its ability to handle novel scenarios people haven’t seen before. That makes it a powerful tool in fields like education, clinical psychology, and user experience research. Instead of running dozens of experiments or surveys, researchers could use Centaur to simulate behavior and explore how people might react to new ideas or policies.

But the real intrigue lies ahead. As the model evolves to include variables like age, personality traits, or socio‑economic backgrounds, it could one day help personalize learning programs, anticipate mental health risks, or guide everything from marketing campaigns to public policy based on how humans actually behave in real life.

Source

This Week’s Conundrum
A difficult problem or question that doesn't have a clear or easy solution.

The Workplace Proxy Agent Conundrum

Early AI proxies can already write updates and handle simple back-and-forth exchanges with human teammates. Soon, they will join calls, resolve small conflicts, and build rapport in your name. Many will see this as a path to increase focus on “real work”, while their proxies handle the routines.

But for many people, showing up is the real work. Presence earns trust, signals respect, and reveals judgment under pressure. When proxies stand in, the people who keep showing up themselves may start looking inefficient, while those who proxy everything may quietly lose the trust that presence once built.

The conundrum
If AI proxies take over the moments where presence earns trust, does showing up become a liability or a privilege?

Do we gain freedom to focus, or lose the human presence that once built careers?

Want to go deeper on this conundrum?
Listen/watch our AI hosted episode

Moon Valley’s Mare AI Filmmaking Tool Goes Public
Moon Valley officially released Mare, its advanced text-to-video model, to the public. Co-founded by former DeepMind researchers, Mare stands out by using only fully licensed training data and giving filmmakers fine-grained control over everything from camera movement to lighting. Upcoming updates will add even more detailed 3D and lighting controls.

Deeper Insight:
By focusing on copyright compliance and filmmaker-level features, Mare aims to break away from the generic “video from text” crowd. As legal battles over AI training heat up, Moon Valley’s approach could give it an edge with professional creators and studios who want both quality and peace of mind.

Grok 4 Released as XAI Pushes Boundaries in Work Culture
XAI debuted Grok 4, its next-generation language model, with a live demo. Reports from inside XAI highlight a relentless work culture with staff sleeping in tents at the office to push the project forward. Grok 4 is a major contender for the most advanced reasoning AI, but recent “improvements” have led to high-profile controversies over offensive or inaccurate content. And there are indications that Grok 4 consults Elon Musk’s X content specifically when answering queries, hmmm. Nonetheless, Grok 4 quickly matched the SOTA benchmarks across the full range of tests, and doubled the prior leader’s score on the Arc-AGI 2 intelligence test, though it only matched the prior top result on the Arc-AGI 1 test.

Deeper Insight:
Grok 4’s launch will create a new conundrum, pitting XAI’s philosophy of open speech in a superlative intelligence against other models with better moderation and principled responses. XAI’s hands-off approach puts pressure on users to filter bad outputs themselves, a stance that could invite regulatory scrutiny as AI adoption widens.

Hugging Face Sells $299 Reachy Mini Open Source Robot
Hugging Face launched the Reachy Mini desktop robot for $299 when connected to the cloud-hosted AI. Pay a bit more $459, and it is powered by a local Raspberry Pi board that can run small AI models. Reachy Mini is fully open source, allowing users to tinker with both hardware and software. It’s aimed at makers, educators, and developers looking to explore embodied AI and experiment with custom language and vision models without necessarily sending data to the cloud.

Deeper Insight:
Open hardware robots are a counterweight to the “black box” bots expected from Big Tech like Tesla and Unitree and Figure. As personal robotics enters the home, open platforms like Reachy Mini could help set early standards for privacy and local AI control.

Meta Poaches Apple’s AI Talent as Siri Lags Behind
Meta continued its talent-poaching, by hiring one of Apple’s top AI researchers, deepening concern about Apple’s stalled progress in bringing a next-generation Siri to market. Apple is in talks to license models from OpenAI or Anthropic, with reports suggesting Anthropic may be in the lead thanks to its safety pitch and infrastructure fit.

Deeper Insight:
As Apple shifts from building to buying its core AI, its privacy-centric approach hinges on running outside models securely on Apple’s own infrastructure. If Meta and others keep pulling away top talent, Apple could struggle to keep up with the next wave of AI-native features.

Wimbledon’s AI Line Judge Draws Player Criticism
Wimbledon adopted an AI-based electronic line calling system for the 2025 tournament, replacing human judges. Players quickly criticized the tech after controversial calls and poor performance in dim lighting, raising doubts about the system’s readiness.

Deeper Insight:
This is a cautionary tale about the risks of over-reliance on AI in high-stakes, real-world settings. Glitches and poor edge-case handling can erode trust fast, especially when the stakes are global and emotions run high.

Researchers “Game” AI Peer Review With Positive Phrases
Researchers from leading universities have started embedding hidden instructions to “focus on positive reviews”, these prompt phrases in academic papers exploit AI-assisted peer review systems. As journals turn to AI to manage the deluge of submissions, some authors are tweaking language to boost their acceptance odds.

Deeper Insight:
Automating peer review brings speed, but also new ways to manipulate the process. This arms race between authors and reviewers shows that quality control will always need some human oversight, at least for now.

Google’s AI Overviews Hit With EU Antitrust Complaint
A coalition of European publishers filed an antitrust complaint against Google’s AI Overviews, claiming the feature diverts traffic, harms revenue, and undermines their businesses by surfacing AI summaries instead of linking to source content.

Deeper Insight:
As search shifts from links to answers, publishers worldwide are confronting a new economic reality. Expect more legal battles, and perhaps new business models, as platforms, regulators, and media companies fight over who controls the flow of information.

Microsoft’s Xbox Advice to RIF’d employees: “Try a Chatbot for Support”
Microsoft’s gaming division was hit by layoffs, and leadership suggested departing employees use AI chatbots to help cope with the transition. The advice, while intended to be supportive, landed as tone-deaf for many losing their jobs.

Deeper Insight:
Corporate reliance on “AI for everything” can easily backfire when empathy and genuine support are needed most. As chatbots become more common, companies must balance automation with authentic human connection.

AI Voice Deepfake Impersonates Senator Rubio in Scam Calls
An AI-generated voice impersonating Senator Marco Rubio was used to contact government officials, including foreign ministers and members of Congress. This case highlights the growing threat of AI-powered impersonation in political and security contexts.

Deeper Insight:
As voice cloning gets easier, public figures and organizations must step up verification and security. This is only the beginning of an era where “who you hear” may not be who you think.

OpenAI, Microsoft, and Anthropic Launch $23M AI Teacher Training Hub
OpenAI, Microsoft, and Anthropic, with the American Federation of Teachers, launched a $23 million initiative to train up to 400,000 educators on integrating AI into classrooms. The program includes resources, model access, and AI literacy efforts.

Deeper Insight:
After months of debate about AI’s risks in schools, this signals a pivot to proactive education. If successful, these partnerships could help shape responsible, creative AI adoption by the next generation of teachers and students.

Consolidation Surges in AI Data Industry
Major AI companies are snapping up data providers at record pace: Databricks acquired Neon for $1B, Salesforce bought Informatica for $8B, and others are eyeing similar moves. The goal: lock up access to proprietary datasets and stay ahead as synthetic data and open access make generic data less valuable.

Deeper Insight:
Data consolidation signals a new phase of competition in AI. As the value of generic internet data drops, the race is on to secure exclusive, high-quality datasets, raising the stakes for both innovation and privacy.

Low-Code Coding Agents Level Up: Lovable, Replit, Emergent Battle for Builders
Lovable launched a new “agent mode” that reduces code build errors by 90%, while Replit added “Dynamic Intelligence” mode with web search and better context awareness. A new tool at Emergent.sh is making waves for its autonomous, full-stack app-building, offering one-click clones of popular SaaS tools, front to back.

Deeper Insight:
No-code and low-code AI coding agents are making it practical for individuals and small teams to build custom tools, often matching enterprise SaaS for internal needs. As these agents improve, companies will rethink what software they buy versus what they can spin up themselves.

CoreWeave Buys Core Scientific, Mistral and LangChain Hit Unicorn Status
CoreWeave acquired data center operator Core Scientific for $9 billion, boosting its AI compute muscle. Meanwhile, both Mistral and LangChain are nearing or surpassing $1 billion valuations, reflecting the ongoing boom in AI infrastructure and tooling.

Deeper Insight:
These moves underscore just how vital compute resources and AI middleware are to the AI stack. Owning the pipes and the platforms that connect models to real-world use cases is the new battleground for AI’s biggest players.

Scientists Pinpoint When AI “Understands” Language
Researchers published findings showing that large language models like GPT shift from analyzing word position to connecting word meaning once they hit a certain training threshold or a “phase change” in language comprehension.

Deeper Insight:
Pinning down when models start to grasp meaning is key for developing smarter, safer AI. This kind of foundational research will shape both the science and the policy of next-generation language models.