‘Her’ AI, Almost Here? Llama 3, Vasa-1, and Altman ‘Plugging Into Everything You Want To Do’

Meta’s newly released Llama 3 70B is arriving in a competitive state—without the full “biggest and best” model or its research paper yet—while Microsoft’s Vasa-1 and Boston Dynamics’ Atlas keep pushing AI toward more lifelike, real-time social interaction. The immediate takeaway: smaller Llama 3 models are already matching top-tier peers on several benchmarks, and the next wave of AI interfaces may be driven less by raw intelligence and more by personalization, avatar realism, and integration into everyday tasks.

Meta’s late-breaking update centers on two smaller Llama 3 releases, with Llama 3 70B positioned as competitive with Gemini Pro 1.5 and Claude Sonnet. Human-evaluated comparisons are presented against models including Mistral Medium, Claude Sonnet, and GPT-3.5, with results described as broadly similar across a “mystery model” still training and newer GPT-4 Turbo and Claude 3 Opus baselines. On graduate-stem style assessment, performance is described as nearly identical, while coding benchmarks show close competition—though the transcript flags that some benchmarks are “deeply flawed,” especially for math where GPT-4 still appears ahead. The bigger research claim tied to Llama 3 is that performance keeps improving even after training on far more data than the “Chinchilla optimal” amount, suggesting saturation from high-quality data rather than sheer volume; coding data receives special emphasis.

Meta also signals a roadmap: multiple future Llama 3 variants with multimodality, multilingual conversation, longer context windows, and stronger overall capabilities. The transcript repeatedly stresses that the most consequential missing piece is not just model size, but the eventual release of the larger model and the paper that will clarify training details like context window length.

In parallel, Microsoft’s Vasa-1 paper spotlights a different kind of progress: generating expressive, controllable facial animation from minimal inputs. The system takes a single image plus an audio clip and produces video with detailed facial dynamics—blinking, lip motion, eyebrow and gaze behavior—at about 40 frames per second with “negligible” starting latency. The method maps facial dynamics into a latent space and uses a diffusion Transformer to connect audio to head movement and facial expression codes before rendering frames using identity features from the input image. Vasa-1 is described as trained on public VoxCeleb2 plus a smaller supplemental dataset, with the transcript emphasizing the surprisingly limited scale (2,000 hours in VoxCeleb2) compared with the massive web-scale data used by frontier systems.

The practical implication is social: Vasa-1 is framed as enabling real-time, lifelike avatar interactions, including in healthcare contexts. The transcript then links this to AI nurse deployments that reportedly outperform human nurses on bedside manner and technical patient education tasks, including identifying medication impacts on lab values and detecting toxic dosages—though the claims are presented as performance metrics and human ratings.

Finally, the transcript ties these advances to a broader debate about personalization versus general intelligence. Sam Altman is cited suggesting long-term differentiation will come from AI models that plug into a user’s life context and integrate across everything they want to do. Meanwhile, AI safety timelines remain contested, with references to Anthropic’s Dario Amodei discussing rapid paths to higher-risk autonomy (ASL 3 and ASL 4). Taken together, the central message is that “Her”-like capability may arrive through tighter personalization and more convincing real-time avatars—potentially sooner than traditional AGI timelines imply.