I have gone through more disease maps in the past six months than in the years before. Claude science or gemini for science opens papers on natural product chemistry, women health and other topics of scientific gaps that occupy my mind that I would not have found alone. One path led to Enveda (I mean, just look at the love in the design and involvement of senses). I am not sure how far ahead this puts me as a scientist, or whether the volume is still useful.
Scientists still say they have to train the model. For the first time the reinforcement learning loop does not need to be coded. The models arrive already pretrained. If someone wants to go further they can build a world model of their own hypotheses and the context around them. The first move is to stop treating every new output as something that must be defended or destroyed. It is just another hypothesis. Producing one is cheap and will only get cheaper. Use as many as can be framed, then box the ones that fail so the next round starts from a cleaner place.
Scientists are being told they have to start using AI on the next paper and the next data set.
2 years ago, I had two options: wait for the formal training programme that keeps getting postponed, or open the bot on the concrete task that is already late.
The effort is specific; the possibilities are open.
Someone can paste a methods section that has sat for three weeks and ask the bot/agent/coworker etc. for the missing controls in plain language. Someone can drop a messy CSV and get a first pass at the cleaning steps, then decide which ones to keep. Someone can feed the bot the rejection letter and ask for the three claims the reviewer actually disagreed with. Maybe the next experiment starts a day earlier. Maybe the figure that was stuck gets drawn. Maybe the bot returns nonsense and the scientist spends an hour fixing it. Those are possibilities, and things I can claim have happened.
Extracting decisions from papers already written
A paper describes PharmAgents2, an autonomous system for end-to-end drug discovery. It extracts transferable decision knowledge from published methods (Paper2Skill), then runs parallel agents that inherit context to explore targets, hypotheses, and strategies at scale. On three real tasks it produced candidates that later showed activity in human-run experiments: a novel LpxH inhibitor scaffold, covalent JAK1 inhibitors with measurable cellular effect, and a new binding target for a known molecule confirmed by SPR. The system still needed wet-lab confirmation. The pattern is the same as a single scientist using the bot on one late task, only scaled.
I can choose to log the exact prompt that cut a literature search from four hours to forty minutes, including the part that still needed a human to verify. I can put the log where another scientist can find it. I can't choose which lab picks it up or whether their PI treats the output as a draft or as finished work.
People who keep moving do not wait for the conditions to become clean or for someone else to unblock them. They treat the obstacle as temporary and keep a plan running. For scientists the obstacle is often "I don't know how to use this yet" or "the tool might invent things" or "the group hasn't decided on a policy." The version that moves is to pick one task that is already late, run it through the bot, record what broke, and do the next task the same way. In the drug-discovery system the parallel agents only worked because each inherited the decision knowledge and the current evidence; a scientist can do the smaller version by giving the bot the previous result and asking for the next check. really small steps of outcome are easy evidence.
Maybe a group starts treating the bot as a colleague who makes mistakes and has to be checked, the same way they treat a new postdoc. Maybe they stop waiting for the university-wide workshop. Maybe they just keep the old workflow because the new one still feels slower on day one. Those are possibilities, not things I can claim have happened.
I can choose to write down what the bot got wrong on a real analysis and leave it on the site. I can't choose whether anyone else runs the same check or whether their field makes verification harder.
Scientists have to start. The choice is whether to wait for the conditions to become clean or to start on the next concrete piece of work and keep a record of what the bot actually does.
The next analysis is waiting, and the bot is already open.