Our Future is WILD! AI Advancements that Get Me EXCITED!

ChatGPT’s new “bring a GPT into the conversation” feature is a meaningful step toward AI assistants that can borrow specialized expertise on demand—though it still feels clunky in how context and discovery work. Users can type an @ mention to pull a custom GPT into an ongoing chat, letting that GPT operate with the full conversation context. In a test, a general question about snail relatives is answered normally, then “Consensus” (an AI research assistant trained on a large academic corpus) is added to verify the claim with citable studies. The workflow works well for quick fact-checking, but it has sharp edges: switching back to regular ChatGPT requires manually dismissing the inserted GPT, and the system tends to treat the added GPT as part of itself rather than a truly separate agent.

That “agent separation” issue shows up in another experiment using “Riz GPT,” a dating-focused assistant. After asking Riz GPT for advice, the user exits it and asks ChatGPT what it just told them; ChatGPT initially responds as if it were the same system. Only after explicitly correcting the context does it respond appropriately. For everyday productivity this may not matter much, but it undermines the promise of modular, multi-agent collaboration. Two other usability gaps stand out: multiple GPTs can’t be stacked at once (they don’t combine if added in sequence), and discovery is limited to recent or pinned GPTs—there’s no search across the broader GPT store for a specific capability when it isn’t already pinned.

Beyond ChatGPT, the most concrete “buildable” development is Meta’s release of Code Llama 70b, positioned as a more performant code-generation model available under the same licenses as earlier Code Llama models. The appeal is practical: code is the interface between humans and computers, and stronger open-source coding models could lower the barrier for creating new tools and fine-tuned variants. Access isn’t a simple download—users must fill out information and are subject to Meta’s discretion—but the open-source framing is central to the excitement.

On the research side, a video enhancement system described with “guided dynamic filtering” and “iterative feature refinement” targets both resolution and motion blur. In examples, low-resolution, fast-motion footage becomes clearer with motion blur “almost entirely solved,” producing sharper faces and more legible text. The presenter argues this could enable higher-quality slow motion without the usual camera tradeoffs (higher shutter speed and frame rates often reduce resolution or strain camera processing). The same theme—improving what cameras can capture—appears again in comparisons where the AI-enhanced results look closer to real footage than other baselines, especially for text and wheel detail.

Other items range from speculative to experimental: claims that large language models have detectable “neural signatures” that differ between truthful and knowingly dishonest behavior; “Morpheus -1,” a multimodal generative ultrasonic Transformer aimed at inducing lucid dreams via neurosimulation rather than word-based prompting; and the idea that future interfaces could extend beyond screens toward brain-state-driven experiences. Taken together, the thread’s throughline is clear: AI is moving from single-chat answers toward tool-using systems, better perception (video), and even direct physiological interaction—while usability and agent boundaries still lag behind the ambition.