AI Reasoning and Crowdfunding Redefine How Science Gets Done

Synapse loss in hibernating mice challenges memory assumptions

A recent study shows that hibernation slashes the number of synapses in mouse brains, yet the animals appear to retain their memories. The finding upends the long‑standing belief that synaptic density directly governs recall ability.

Researchers observed that during hibernation mice lose a substantial fraction of synaptic connections, but behavioral tests indicated that the rodents still navigate mazes they learned before winter. The paper, published in Ars Technica’s science roundup, notes the paradox without offering a mechanistic explanation, leaving a gap that computational models could help fill.

The result matters because it suggests the brain can preserve functional circuits despite massive structural pruning. If AI reasoning systems can model such resilience, they might inspire new algorithms for fault‑tolerant computing.

Crowdfunding science: Petridish’s model and early traction

Petridish launched as a niche crowdfunding platform that connects science enthusiasts directly with researchers seeking project funds. Co‑founders Ilia Papas and Matt Salzberg—formerly a senior associate at Bessemer Venture Partners—designed the service to give the public a say in which experiments get off the ground.

On the site, scientists post project briefs, and backers receive insider updates, naming rights, or field‑trip invitations in exchange for contributions. Petridish takes a flat 5 percent of each donation, and projects only receive money if they meet their funding goal before the deadline.

Current campaigns span a wide range: decoding gelada monkey melodies, capturing deep‑sea creature sounds, and cataloguing new ant species in Madagascar. One researcher even offers naming rights to any new animal discovered, a perk that has attracted hobbyists eager for a slice of scientific legacy.

The model sidesteps equity expectations common on platforms like Kickstarter, focusing instead on intellectual rewards. By putting decision‑making power in the hands of a broader community, Petridish hopes to diversify the kinds of research that attract financing.

Gemini 3 Deep Think: AI reasoning enters the lab bench

Google unveiled a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode built for science, research, and engineering challenges. The new version is now available to Google AI Ultra subscribers via the Gemini app and, for the first time, via the Gemini API to select researchers, engineers, and enterprises.

Early testers report concrete wins. Lisa Carbone, a mathematician at Rutgers University, fed Deep Think a highly technical physics paper and the system flagged a subtle logical flaw that human peer reviewers missed. At Duke University, the Wang Lab used Deep Think to design a crystal‑growth recipe that produced thin films over 100 µm thick—precisely the target that prior methods struggled to hit.

Google’s R&D lead Anupam Pathak also leveraged Deep Think to accelerate component design, though the report does not detail the specific parts. Performance metrics show the upgraded model achieving gold‑medal‑level scores on the written sections of the 2025 International Physics Olympiad and Chemistry Olympiad, and a 50.5 % score on the CMT‑Benchmark for advanced theoretical physics.

Deep Think’s strength lies in handling messy, incomplete data and offering practical suggestions rather than abstract theory. The system’s ability to spot errors in peer‑reviewed literature hints at a future where AI assistants become routine co‑authors on scientific papers.

The convergence: New funding + AI tools reshape research workflow

When a platform like Petridish democratizes funding and an AI engine like Deep Think streamlines analysis, the research pipeline contracts dramatically. A scientist can now pitch a project to a global audience, secure micro‑funding, and run preliminary simulations through an AI model before any wet‑lab work begins.

Consider a hypothetical study on synaptic pruning during hibernation. A researcher could launch a Petridish campaign, offering backers naming rights to any novel neural pathway discovered. Simultaneously, the team could feed electrophysiology data into Deep Think to generate hypotheses about how memory persists despite synapse loss. The AI’s suggestions could guide experiment design, reducing trial‑and‑error cycles and making the crowdfunded budget stretch further.

Critics warn that reliance on AI could embed hidden biases, and that crowd‑sourced funding might favor flashy projects over incremental science. The Petridish model mitigates some of that risk by requiring projects to meet concrete funding thresholds, but the platform still depends on public interest, which can be fickle. Meanwhile, Deep Think’s early successes are promising, yet the system’s “gold‑medal” benchmarks do not guarantee reproducibility in real‑world labs.

The real test will be whether these tools can sustain long‑term, high‑impact research without sacrificing rigor. If they do, the traditional gatekeepers—large grant agencies and elite journals—may find their influence waning.

What to watch

Track the next round of Deep Think API access requests; Google plans to open the service to a broader set of institutions later this year. On the funding side, monitor Petridish’s campaign success rates and any policy shifts around donor attribution, especially as naming‑rights incentives become more common. Finally, watch for follow‑up studies on hibernating mice that use AI‑driven analysis to pinpoint the molecular mechanisms preserving memory despite synapse loss. Those three threads will reveal whether the hype translates into a durable shift in how science is financed and executed.