Nate Jones and Shira Lazar Review 2025 in AI

In a December 2025 year-end episode, Shira Lazar and analyst Nate Jones look back at major AI stories and offer predictions for 2026. Their assessments and forecasts reflect the time of recording.

They begin with GPT-5.2. Jones says the improvements he finds most interesting concern demanding workflows, such as processing a large spreadsheet and producing a presentation, rather than ordinary chatbot conversation.

Jones predicts that casual chatbot use may feel less different between model releases, while more advanced workflows continue to change. This is his assessment in 2025, not a conclusion demonstrated by the episode.

The conversation turns to competition with Google’s Gemini 3. Jones interprets the competitive pressure as important to OpenAI’s funding and infrastructure strategy; he presents that explanation as his view.

Shira highlights Gemini features and argues that broad model capabilities can affect smaller single-purpose products. Jones expects that trend to continue as foundation models take on more types of tasks.

Discussing Claude Opus 4.5, Jones says non-programmers can increasingly start building websites or applications with AI. He also points out that translating intent into a working product is still difficult.

Unexpected questions about deployment, hosting and other software details can appear after a first prototype. Jones recommends learning enough technical concepts to communicate and make decisions, even without becoming a professional programmer.

Shira discusses DeepSeek’s impact on the 2025 model landscape and the attention paid to its reported cost. Jones argues that widely repeated cost headlines omitted important context; his interpretation of timing and motives is speculative.

They discuss the difference between a narrow benchmark result and performance on more general tasks. Jones says some models can be brittle and argues that more consistent performance is an important area for improvement.

Jones sees open models as a route to making AI available across many devices, even when a local model is not the strongest available. Shira notes that she follows the topic without personally working deeply with open models.

The pair turn to Sora and Hollywood. They discuss a Disney and OpenAI agreement and possible reasons a studio might license characters. Jones’s account of private negotiations is conjecture, not something established in the conversation.

They ask how performers and unions can respond as synthetic characters and licensing arrangements develop. Shira says the pace of agreements raises questions about how creative workers keep up.

On copyright disputes, the speakers discuss litigation and a settlement involving authors and Anthropic. Their comments are a 2025 discussion of particular proceedings, not a comprehensive statement of copyright law or the status of later cases.

They debate how transformative use, licensed material and settlements may shape the industry. Both speculate about why companies chose particular strategies; the conversation does not resolve the underlying legal questions.

The discussion moves to AI music. Shira cites examples of generated voices and licensing, while Jones distinguishes short-lived streaming hits from the live performances through which many artists build careers.

Jones expects some artists to use AI for components of studio production while others continue to value recognizable human collaborators. Shira compares the debate, cautiously, to earlier arguments over production technologies.

They discuss a synthetic performer and speculate about where generated characters might appear in film. Jones distinguishes productions built around recognizable stars from genres where audiences may care less about individual performers.

Shira raises disclosure of AI use in awards and film production. Jones describes seeing an AI-generated version of himself speaking another language and imagines localization uses, while Shira raises possible effects on jobs.

On image generation, both praise Nano Banana as of the recording. Jones predicts that competition will continue; their ranking is a dated opinion, not a current comparison.

Shira asks whether image tools contribute to “AI slop.” Jones argues that low-quality work predates AI and suggests examining both the volume of generated material and standards applied to human work.

Their final 2025 topic is data centers and sustainability. Jones says water and electricity should be considered separately and offers comparisons with other activities. The episode does not supply enough detail to verify those comparisons or generalize them across locations.

Jones argues for comparing AI with other industries and says the public can ask for more renewable electricity. He acknowledges that data-center projects can rely on natural gas while trying to meet rapid buildout deadlines.

They discuss siting renewable infrastructure and the difficulty of local opposition. Jones offers his interpretation of China’s buildout; these are the speakers’ views rather than an environmental assessment.

Looking ahead from December 2025, they discuss possible US AI regulation. Jones predicts limited federal action except perhaps on child protection, and they consider potential state and international approaches. These are historical forecasts, not a statement of current law.

On text-to-video, Jones expects advertising platforms to automate more creative production and predicts an increase in synthetic ads. He also raises the open question of whether audiences will form durable relationships with synthetic personas.

Shira says her concern is less a single synthetic influencer than a crowded information environment. She plans to emphasize offline events, expertise and relationships as ways to build trust.

Jones agrees that original thinking and in-person gatherings may matter more when audiences are uncertain what is real. Both treat this as a prediction about behavior rather than a proven outcome.

They discuss politics, deepfakes and trust. Jones argues that confirmation bias predates AI, while Shira emphasizes the sophistication of coordinated misinformation and the need to question what people consume and share.

Returning to sustainability, Jones hopes for more precise discussion of local impacts, including water temperature and ecosystems. Shira says clear expectations could help identify opportunities for lower-impact approaches.

On education, they hope AI might support individualized learning while allowing teachers to focus on teaching. Jones imagines a system that explains concepts to a student instead of simply supplying answers.

Shira sees gaps in current tools and policy as opportunities for people to build solutions. Jones closes by saying the future need not be terrifying, but keeping up requires sustained, intentional effort.

Shira thanks Jones, points listeners toward his work and closes the year-end episode. The episode’s later sponsor and channel announcements are omitted from the recap.