Research frontiers.

AI-designed drugs reach the clinic

Isomorphic Labs (Alphabet)

The first real-world test of AI drug discovery: Isomorphic (17 programmes, oncology/immunology, with Lilly/Novartis/J&J) targets first-in-human trials by end of 2026 — Hassabis's on-record Davos statement, itself a one-year slip.

As of 30 Sep 2026 the company had not announced an IND filing or trial start, so the date is live but its slipping risk is documented.

• 17 programmes with Lilly/Novartis/J&J; first-in-human target end of 2026 (source)
• As of 30 Sep 2026, no IND filing or trial start announced (source)

Dates are company targets; this programme has already slipped once.

Checked Oct 14, 2026

Brain-computer interfaces (AI-decoded)

Neuralink

Per Neuralink's published roadmap (company claims — labelled as such): Blindsight first human trial targeted 2026, multi-region/multi-device implants 2027, scale to 25,000 electrodes and AI-integration work 2028.

21 patients implanted as of recent reporting; high-volume production and automated surgery are the 2026 programme.

• Roadmap: Blindsight trial targeted 2026; 25,000 electrodes and AI integration by 2028 (source)
• 21 patients implanted as of recent reporting (source)

Every date here is a company target — this company's timelines have slipped before.

Checked Oct 14, 2026

Continual learning

Academia + labs (no owner)

The blocker has a name — the 'Memento Constraint': models that learn continuously without catastrophic forgetting. A May 2026 research map finds forgetting rates of 40-80% under standard fine-tuning and nothing production-ready; first genuinely continual frontier models are

estimated 2028-2030. This is the frontier most likely to change what a 'model release' even means — and it is not close.

• Forgetting rates of 40-80% under standard fine-tuning; five competing approach families (source)
• First genuinely continual frontier models estimated 2028-2030 (source)

Checked Oct 14, 2026

Embodied AI / robot foundation models

Google DeepMind; Figure AI; Tesla

The race to one model that perceives, plans and actuates across robot bodies: DeepMind's Gemini Robotics line (ER-1.6 shipped Apr 2026) and Figure's Helix are the visible programmes, consolidating video, 3D and control into single physical foundation models.

Progress is real but gated by hardware production, not just models.

• Gemini Robotics ER-1.6 shipped Apr 2026 (source)
• Figure's Helix programme is backed by a $3.5B compute commitment (source)

Checked Oct 14, 2026

World models

Google DeepMind; World Labs (to AMD)

Genie 3 generates real-time, interactive, physics-consistent worlds from prompts, and DeepMind presents world models as robot-training infrastructure, not just media.

World Labs is being acquired by AMD for ~$8.2B in stock (announced 28 Sep 2026, expected to close by end of 2026); the frontier to watch is persistent, simulation-grade worlds used to train robots.

• Genie 3 generates interactive, physics-consistent worlds from prompts (source)
• AMD acquiring World Labs for ~$8.2B in stock, announced 28 Sep 2026 (source)

Checked Oct 14, 2026