Rohit Krishnan argues that reasoning models should let users pause, inspect, and redirect their intermediate steps instead of forcing all instructions into the initial prompt.
2
Rohit Krishnan expects AI to change ordinary jobs by turning workers into sources of process data that can later train systems to automate their work.
3
Rohit Krishnan grounds AGI predictions in physical limits such as chips, power, failure rates, manufacturing capacity, and the cost of running agents continuously.
Summary
Rohit Krishnan and Demetrios Brinkmann discuss how people should interact with reasoning models, especially systems that produce long reports after searching and thinking for several minutes. Rohit argues that users should be able to pause a model, add context, ask it to search further, or branch from an interesting reasoning step. They then discuss how AI may change jobs. Workers may increasingly generate the process data needed to automate parts of their own work, while some roles could disappear or change into orchestration work. Rohit treats the resulting increase in AI output as a possible source of useful velocity, although he says fields such as academic publishing could struggle to review the extra volume. On AGI, he rejects predictions that ignore physical constraints. The number of chips, their power use, service life, failure rates, and manufacturing capacity all shape what can run. He also explains his work at Bodo.AI, which makes large-scale data engineering in Python faster and has moved into open source.
Reasoning models should let users take over at an interesting step
Rohit Krishnan sees reasoning models as a new interaction problem. A chatbot gives an answer, while a tool-using model can be stopped when it chooses the wrong tool. A reasoning model exposes a sequence of intermediate steps, so users should be able to pause when they see a useful idea and continue from there. For a scientific question, that might mean searching for evidence or writing code to test an assumption. For literature, it could mean developing an unexpected connection instead of letting the model move past it. Rohit compares this to checkpointing. The user can add context, ask for more depth, or start another action from the same point.
Long reports need feedback during generation, not only after delivery
Rohit Krishnan says the usual workflow for a long AI report is inefficient. OpenAI's deep research feature asks questions, refines a plan, and works for 10, 15, or 30 minutes. Its output can be strong, but the most interesting material may be buried deep in the process. Rohit would rather see the reasoning as it happens and leave comments that trigger more work on a specific section. Users could ask the model to search further in one area instead of receiving thousands of words and then trying to repair the result afterward. This would make report creation more like collaboration.
Adding context while a model works could reduce the user's cognitive burden
Demetrios Brinkmann frames the problem as cognitive load. If a user must provide every detail before a model starts, wait for a long response, inspect it, and repeat the cycle, the process becomes expensive to manage. Rohit thinks pausing is technically feasible because models already produce multiple possible reasoning rollouts and can preserve what happened before. A system could offer common actions such as 'go deeper here' or 'search the web for this'. Tool use makes the idea more natural, since models already pause to write code, search documents, or retrieve outside information. The user's input does not have to be front-loaded.
AI may turn much of ordinary work into training data for future automation
Rohit Krishnan expects models to become useful for real work, such as understanding a codebase and contributing a pull request. That raises a socioeconomic question. Analysts, researchers, and other roles may change substantially or disappear as systems learn from the work people perform. He extends Demetrios Brinkmann's point about data labelers: workers doing high-value tasks may also become training wheels for future AI because their processes and decisions can be captured. A person could do a job while the surrounding system records enough examples to automate parts of it later. Rohit does not know whether the result will be more orchestration work or a deeper change in the labor market, but he sees the mechanism as plausible.
More AI output can make noise acceptable when work becomes cheap to redo
Demetrios Brinkmann worries that AI will flood teams with reports and pull requests. Rohit agrees that many AI-generated pull requests may be poor, but says another model can review them and that the work may become cheap enough to repeat. In that setting, pull requests are commoditized. Their higher volume and variation may be tolerable if people can still find useful contributions, much as they find signal in a noisy social platform. He draws a sharper limit around academia. If researchers move from publishing monthly to publishing weekly, journals and reviewers may not keep up. The result would challenge the existing publishing system, much as faster software release cycles changed older release hierarchies.
AGI forecasts need to account for chips, energy, and manufacturing
Rohit Krishnan wrote about AGI by starting with a base case rather than assuming unlimited resources or recursive self-improvement. He asks what happens if a human-level agent needs a continuously running H100, then considers the number of GPUs, their power consumption, service life, utilization, and failure rates. He gives an illustrative estimate of roughly 40 to 45 million full-time agents under those assumptions, while stressing that the estimate is not the main point. To reach far larger numbers, the assumptions about efficiency, hardware, or energy must change. Manufacturing capacity also matters. ASML, TSMC, and chip suppliers cannot expand without limits, so physical bottlenecks move as technology improves instead of vanishing.
Rohit Krishnan describes Bodo.AI as a way to run data engineering and analytics on large datasets in Python. It targets transformations, processing, and preprocessing that teams might otherwise handle with Spark, PySpark, or Dask. Bodo.AI recently moved into open source, and Rohit wants users to try the examples, inspect the benchmarks, and find where the system breaks. He says the product can make pandas, NumPy, and Polars workloads much faster, with cited benchmark differences ranging from 20 to 250 times depending on the comparison. He connects this work to AI because raw data must be cleaned, processed, and prepared before training or inference. The company is also exploring parallel inference and AI-readable documentation that could help models make code Bodo compatible.
"The goal is we want to get as many people as possible to try and play with it, use it, get an understanding from the examples that we have already put out there, see the benchmarks that we put out there, knock on it, figure out where it breaks."Rohit Krishnan44:33
Who should watch
You are designing an interface for reasoning or research agents and need concrete ideas for pausing, branching, and adding context during a run.
Your AI plans assume that more agents automatically mean more productivity, and you want a discussion grounded in hardware, energy, and manufacturing limits.
You work on data preparation for machine learning and want to hear why Rohit Krishnan built Bodo.AI around faster Python data processing.
