Interesting - I’m a researcher in renewable energy and our company is now using Claude. Previously I used Perplexity - the free version. After one week using Claude, I can’t stand it; it doesn’t make work faster, it makes it slower and more confusing. Ask Claude a very specific question with guardrails and it will still throw in the kitchen sink and the dog’s dinner into the answer. It comes up with stuff that isn’t true, it exaggerates wildly. I’ve gone back to Perplexity; it isn’t always totally honest but it is so much better for straightforward research where it gives you all the sources and will answer questions in a direct manner.
Hmmm, no I haven’t adjusted the settings yet. I will check them now though - thanks for pointing that out. I’m a bit stuck in trying to create documents that can be understood at a glance or in a very high level. The problem is that renewable energy is extremely complex. The landscape, the technology, regional regulations, money (revenue and investment), geo politics…so it is very difficult to create answers in a simple way. My answers must be backed up by evidence - which effectively Is the granular detail that nobody has the time to read! After reading yours and others comments I’m reorganising how I do my research and how I put it together afterwards
The challenge is building new research processes that can harness near-infinite knowledge capture; combining unprecedented breadth, depth, speed, and quantity without collapsing into chaos.
The risk that scale and acceleration begin to displace the foundations of good research is real. Tools and processes that augment clear questions, purpose, conceptual clarity, critical thinking, critical distance, incubation, reflection, and the productive friction of sustained thought. Serendipity important also. Not all insight emerges through optimisation.
The emerging skill is not simply accessing information, but imposing constraints upon abundance; orchestrating infinite flows of data in service of coherent knowledge production. Curation, judgement, analysis, reasoning, synthesis, and methodological discipline become even more important.
The central question is how to use AI to augment human inquiry without eroding coherence, integrity, and intellectual agency.
The answer (as you articulate above Stephen) is that there is still much to test, negotiate, and learn as these practices evolve.
I would add that the false binary of AI as either universally transformative or universally corrosive is not especially helpful to researchers. The challenge is determining where and how these systems meaningfully contribute to the production of knowledge, and where human judgement, constraints, deliberation, and desirable friction remain essential.
The pollution of publications by AI slop is a problem in every domain. The Jevons paradox problem is spot on - I'm seeing the increase in effort it takes to verify output (ironically as hallucinations get less common, they are harder to spot). As agentic research becomes more mainstream, controlling context will be paramount.
I've been testing out Liminary (www.liminary.io) - could be worth a look (full disclosure - I'm doing some work for them right now). It saves papers, notes, and sources as you go, and it bases its responses in that material rather than hallucinating from training data.
Good question. NotebookLM works well for querying a specific set of documents in a session. Liminary is different in that it builds a persistent knowledge layer over time. Everything you save - meeting transcripts, web pages, uploaded PDFs, videos etc - accumulates and stays available across all your work. So when you're drafting in Google Docs weeks later (or interacting with Liminary in a Google Meet call), it can surface something relevant from a meeting you had a month ago without you asking for it. It's more ambient and ongoing than a session-based tool.
Interested to hear your opinion on Andrej Karpathy's Autoresearch and how a model like that may fit into research spaces other than AI training and coding optimisation.
I am experimenting with this right now—the domain is not really right for economics but some stuff is going to be very interesting indeed, especially now he’s at Anthropic.
Interesting - I’m a researcher in renewable energy and our company is now using Claude. Previously I used Perplexity - the free version. After one week using Claude, I can’t stand it; it doesn’t make work faster, it makes it slower and more confusing. Ask Claude a very specific question with guardrails and it will still throw in the kitchen sink and the dog’s dinner into the answer. It comes up with stuff that isn’t true, it exaggerates wildly. I’ve gone back to Perplexity; it isn’t always totally honest but it is so much better for straightforward research where it gives you all the sources and will answer questions in a direct manner.
Fascinating Geraldine, I must check out perplexity. Have you customised the Claude.md settings?
Hmmm, no I haven’t adjusted the settings yet. I will check them now though - thanks for pointing that out. I’m a bit stuck in trying to create documents that can be understood at a glance or in a very high level. The problem is that renewable energy is extremely complex. The landscape, the technology, regional regulations, money (revenue and investment), geo politics…so it is very difficult to create answers in a simple way. My answers must be backed up by evidence - which effectively Is the granular detail that nobody has the time to read! After reading yours and others comments I’m reorganising how I do my research and how I put it together afterwards
The challenge is building new research processes that can harness near-infinite knowledge capture; combining unprecedented breadth, depth, speed, and quantity without collapsing into chaos.
The risk that scale and acceleration begin to displace the foundations of good research is real. Tools and processes that augment clear questions, purpose, conceptual clarity, critical thinking, critical distance, incubation, reflection, and the productive friction of sustained thought. Serendipity important also. Not all insight emerges through optimisation.
The emerging skill is not simply accessing information, but imposing constraints upon abundance; orchestrating infinite flows of data in service of coherent knowledge production. Curation, judgement, analysis, reasoning, synthesis, and methodological discipline become even more important.
The central question is how to use AI to augment human inquiry without eroding coherence, integrity, and intellectual agency.
The answer (as you articulate above Stephen) is that there is still much to test, negotiate, and learn as these practices evolve.
I would add that the false binary of AI as either universally transformative or universally corrosive is not especially helpful to researchers. The challenge is determining where and how these systems meaningfully contribute to the production of knowledge, and where human judgement, constraints, deliberation, and desirable friction remain essential.
The pollution of publications by AI slop is a problem in every domain. The Jevons paradox problem is spot on - I'm seeing the increase in effort it takes to verify output (ironically as hallucinations get less common, they are harder to spot). As agentic research becomes more mainstream, controlling context will be paramount.
I've been testing out Liminary (www.liminary.io) - could be worth a look (full disclosure - I'm doing some work for them right now). It saves papers, notes, and sources as you go, and it bases its responses in that material rather than hallucinating from training data.
Cool, will check it out Kevin! The reproducible workflow part is why I think Zerve is a great tool, would recommend you taking a look.
Cheers - Zerve looks really interesting and seems to solve an adjacent problem to Liminary actually for the analytics use case.
Interested in Liminary. Is it like a more generative version of Notebooklm - less of scrapper and indexer?
Good question. NotebookLM works well for querying a specific set of documents in a session. Liminary is different in that it builds a persistent knowledge layer over time. Everything you save - meeting transcripts, web pages, uploaded PDFs, videos etc - accumulates and stays available across all your work. So when you're drafting in Google Docs weeks later (or interacting with Liminary in a Google Meet call), it can surface something relevant from a meeting you had a month ago without you asking for it. It's more ambient and ongoing than a session-based tool.
Interested to hear your opinion on Andrej Karpathy's Autoresearch and how a model like that may fit into research spaces other than AI training and coding optimisation.
I am experimenting with this right now—the domain is not really right for economics but some stuff is going to be very interesting indeed, especially now he’s at Anthropic.