How is AI impacting science?

Original Article ↗ Hacker News Discussion ↗

Advances in AI are reshaping scientific work, from high‑profile breakthroughs like AlphaFold’s protein-structure predictions and DeepMind’s algorithmic discovery tools to more mundane but widespread uses such as code assistance and paper summarization. Commenters weigh AI’s potential to deliver “good enough” solutions to hard optimization and simulation problems—sometimes encroaching on the perceived niche of quantum computing—against its current limitations in areas like full biological function prediction. Alongside optimism about AI‑accelerated research, there is growing concern over ethical risks, job displacement, environmental impacts, and a surge of plausible but low‑quality, AI‑generated academic papers that strain peer review.

AI vs. Quantum Computing

  • Some physicists worry that rapid AI progress and “good enough” heuristic solutions to NP-class optimization problems (e.g., logistics, chemistry) could undercut funding for slower-moving quantum computing (QC).
  • Others argue QC’s main advantage is not NP-complete problems but the BQP class (e.g., factoring) and especially simulating quantum systems.
  • There is confusion and debate over whether QC is “massively parallel”; several comments stress that quantum algorithms exploit amplitudes and entanglement rather than classical-style parallelism.
  • Claims that QC can simulate “any quantum effect in O(1)” are challenged as oversimplified and limited by hardware primitives and readout constraints.

Protein Folding, AlphaFold, and Biological Impact

  • AlphaFold is widely seen as a revolutionary advance in structure prediction, dramatically speeding work, but is not a full “solution” to protein folding:
    • It struggles with missense variants and doesn’t directly yield function, catalysis, dynamics, or cellular context.
    • Static structure is only one step in complex pipelines for drug design or understanding mutations.
  • Even perfect primary→quaternary prediction would not by itself transform society; major gaps remain in modeling interactions, dynamics, and whole-cell behavior.
  • Some foresee eventual AI systems that map mutation→functional change and, combined with CRISPR, enable broad genetic therapies, but this is considered far future.
  • Quantum or advanced simulators for whole-cell molecular dynamics are suggested as a next frontier.

Ethics and Limits of Powerful Technology

  • One side argues some knowledge (e.g., protein design, advanced AI) is “too much power” for a short-sighted, unaccountable society; sees AI as net harmful and questions medical and agricultural advances that enabled large populations and ecological damage.
  • Others counter that technology (antibiotics, vaccines, fertilizers) has huge benefits, population growth is slowing, and the core problem is human behavior, governance, and mental health, not technology per se.

LLMs in Scientific Practice

  • Many expect the biggest near-term impact from mundane uses:
    • Helping non-programmer researchers write and debug experiment code.
    • Assisting with reading, summarizing, and writing jargon-heavy papers, especially across fields or languages.
  • Concerns:
    • Researchers may not understand AI-generated code well enough to detect subtle errors, but commenters note this is already true for much academic code.
    • Verification still requires good test design and some coding literacy; AI is a productivity aid, not a substitute for understanding.

Junk Papers and Scientific Publishing

  • Conferences are seeing a surge of plausible-looking but empty, LLM-written submissions, straining reviewer capacity.
  • Proposed responses include pre-review filters and requiring established community members to “vouch” for submissions.
  • Some note that LLM prose has recognizable stylistic tics but these can be reduced with careful prompting, suggesting style-based detection is fragile; calls for explicit disclosure of AI assistance.