Each week there’s no less than one new announcement about what AI will do to software program improvement. A lot of the dialog strikes quick and stays shallow, specializing in productiveness good points, job displacement, or which mannequin scored highest on the newest benchmark.
We’re extra within the questions beneath that noise, so we created the JetBrains Analysis Podcast. In it, we discover questions, similar to:
- What does it imply to grasp a system, not simply generate its code?
- What does psychology inform us about why groups succeed or fail?
- Why does verifiability matter greater than status in coaching brokers?
In our podcast, we’re primarily excited about analysis in its varied varieties and the individuals behind it. Every podcast episode goes someplace attention-grabbing with somebody who has spent severe time on issues that matter. On this weblog submit, we spotlight the concepts that got here up in these conversations.
Cat Hicks: Tradition is infrastructure
“Tradition is infrastructure. It isn’t only a “good to have”. It shapes our considering. It shapes our problem-solving. It immediately adjustments the standard of labor that we do.”
In our newest episode, we communicate with Cat Hicks, psychologist and writer of Psychology of Software program Groups. We ask her: What really determines whether or not software program groups do good work? Why examine groups as an alternative of particular person builders?
Hicks has discovered that indicators of belonging, studying, and recognition from a group can reduce builders’ measured AI id menace, i.e. the concern that their experience is turning into out of date, roughly in half. She emphasizes that you will need to have a look at the group degree to raised perceive what’s going on with particular person group members.
For instance, Hicks has developed a measure she calls overproduction strain, which refers back to the feeling that arises when plainly solely short-term output issues. When that strain is excessive, individuals generate code with out understanding it, mistrust colleagues, and cease being sincere with managers.
Hicks additionally challenges the “lone genius” fantasy: the concept groups ought to be constructed round one highly-talented particular person. As an alternative she argues that it’s really higher to attempt for “extra bands, fewer rock stars”. Extremely collaborative cultures, similar to open-source problem-solving, might be particularly efficient at attaining expertise breakthroughs.
Tomáš Petříček: 5 cultures, one argument
“Programming has by no means actually [been] simply implementing a specification.”
In our fourth episode, we discuss to Tomáš Petříček, a professor at Charles College in Prague who research the historical past and philosophy of software program. His guide Cultures of Programming argues that programming has by no means been a single self-discipline. As an alternative, the next 5 distinct cultures have been competing and collaborating because the earliest days of computing:
- Mathematical, which treats packages as formal objects to be proved appropriate.
- Hacker, which values direct engagement with a working machine.
- Managerial, which is concentrated on predictability and labor group.
- Engineering, which accepts that techniques are advanced and asks which practices make them dependable regardless of that complexity.
- Humanistic, which begins with the query of what occurs in an individual’s thoughts after they work together with software program.
Every tradition has a special thought of what a program is, what good work seems like, and what issues most. These cultures clarify why the identical occasion can result in various diagnoses. Take, for instance, the Knight Capital failure in 2012, the place a software program deployment error value the corporate $460 million in lower than an hour. Relying on which tradition you strategy the issue from, one analysis would possibly differ wildly from one other. Listed here are some instance diagnoses from every of the 5 cultures:
- A mathematician would ask what might have been formally assured.
- A hacker would ask what low-level data would have caught th eproblem.
- An engineer would ask about testing and rollback.
- A supervisor would ask about processes and accountability.
- A humanist would ask whether or not the system made its state legible to these working it.
Petříček additionally talks concerning the present enthusiasm for AI brokers, particularly specification-driven improvement. He emphasizes the significance of historical past. For the reason that Nineteen Sixties, individuals have been making an attempt to isolate the mechanical a part of programming and automate it. Nevertheless it hasn’t labored but, as a result of software program improvement has all the time been extra than simply the implementation of a specification. Groups be taught what they’re constructing via prototypes, conversations, and use.
Alexander Kulikov: What you’re constructing when the answer is unknown
“While you discover an issue that doesn’t get out of your head, that is in all probability the appropriate downside.”
In our third episode, we chat with Alexander Kulikov, who heads the Algorithms and Complexity Principle Lab at JetBrains Analysis and is the Head of Laptop Science and Synthetic Intelligence B.Sc. program at Neapolis College Pafos. His description of the present local weather is extra sincere than most: “superb, thrilling, and scary on the identical time.”
And the priority isn’t fully theoretical. Chess engines, for instance, surpassed each human participant a long time in the past. Kulikov doesn’t know of any legislation stopping AI from finally doing the identical to theorem proving. He doesn’t count on it imminently, however he additionally didn’t count on AI-assisted coding to progress as quick because it has. His place is what he calls attentive uncertainty: neither denial nor resignation.
Within the episode, we additionally discuss what it really means to show pc science in an period when AI can clear up Olympiad issues. Kulikov’s curriculum technique is straightforward: train graduate college students be taught troublesome issues effectively. The precise applied sciences they’ll want are unknowable. The flexibility to dive into an unfamiliar topic, construct a psychological mannequin, and make progress is extra sturdy than any specific talent.
That’s additionally why he nonetheless takes arithmetic critically in a world of AI-assisted coding. A pupil who understands the construction of an issue can check an AI-generated reply, establish lacking assumptions, and get better when the primary try fails. A pupil who solely is aware of request a solution might not know whether or not the reply is reliable.
Kulikov additionally attracts a distinction between researchers who’re “birds” (i.e. ranging throughout fields, recognizing connections from above) and “frogs” (i.e. digging deep into one space). On this analogy, he can be a frog. However AI, he thinks, might act as a man-made hen for specialists: reducing the price of wanting past your fast area with out requiring you to change into a generalist.
Ibragim Badertdinov: From dentistry to coding brokers
“Vibe checks don’t scale.”
In our second episode, we discuss with Ibragim Badertdinov, the Lead Analysis Engineer at Nebius. He took a considerably uncommon path to AI analysis: he first graduated from medical college with honors, accomplished a dental residency, after which gave himself a yr to strive one thing else. On this episode, he talks about how he made that swap, what he realized at College 21 and on Kaggle, and the work that adopted. Most notably, we discuss his work on constructing SWE-rebench.
SWE-rebench is an automatic pipeline for evaluating coding brokers on actual software program engineering duties. Accepted to NeurIPS 2025, it additionally has over 12 million Hugging Face downloads and one million leaderboard visits a month. Its key design characteristic is decontamination, or pulling duties from GitHub solely after a mannequin’s launch date, so brokers can’t be educated on the solutions. This manner, SWE-rebench isolates the mannequin’s problem-solving capacity from any potential publicity to coaching information. After its launch, it made headlines when many fashions carried out considerably worse on SWE-rebench duties than on SWE-bench Verified ones (see the paper for extra particulars).
That is particularly essential, as a result of the usual to this point has been SWE-bench Verified, and the reported scores from this benchmark carried vital weight, each in analysis and the broader trade. As was mentioned in a current weblog submit, this tendency to deal with solely SWE-bench Verified scores may give us a deceptive image of mannequin efficiency. The submit additionally suggests methods to judge fashions extra precisely.
Badertdinov additionally talks about why reinforcement studying with verifiable rewards has pushed such dramatic good points in coding and math particularly. The perception is deceptively easy: in domains the place you may confirm whether or not a solution is appropriate, you may generate huge coaching indicators mechanically. Code runs or it doesn’t; exams move or they don’t. That’s what makes software program engineering duties so invaluable for post-training: not their status, however their verifiability.
Anna Kogan: The library that grew to become infrastructure
“Proper now, pc imaginative and prescient is just not cool. LLMs are very cool.”
Our pilot episode is a dialog with Anna Kogan, founder and CEO of FitWise and former CEO of OpenCV.ai. She’s been in pc imaginative and prescient for 15 years and nonetheless believes in its scientific worth: in pc imaginative and prescient, researchers collect a dataset, practice a mannequin, measure metrics, and clear up particular duties. In our dialog, we talk about challenges within the area, together with monetizing open-source instruments and sluggish adoption of AI in varied industries regardless of the expertise already current.
The hole between affect and resourcing for open-source instruments is one thread within the dialog. An instance we talk about is OpenCV, the open-source pc imaginative and prescient library she helped construct into a worldwide normal. It runs on billions of gadgets and will get 32 million Python downloads a month on PyPI alone. Due to these numbers, many individuals assume the library is maintained by a number of hundred builders, whereas in actuality it’s been solely two to eight builders at any given time.
One other dialog thread considerations a more durable query: Why does pc imaginative and prescient adoption nonetheless transfer so slowly, even when instruments similar to OpenCV are clearly very fashionable? In response, Kogan brings up human motion in sports activities. Although the technological advances are far sufficient alongside that folks in sports activities might cease manually clicking via and annotating broadcast footage, organizations are proof against adopting AI assist. One motive is that it’s nonetheless cheaper to rent employees within the Philippines to do that guide work, than it’s to determine an AI system to take it over.
With FitWise AI, Kogan is excited about serving to sports activities organizations see the worth in pc imaginative and prescient adoption. They supply the information, 3D digital replicas of athletes’ physique shapes and actions in actual time; this information is collected by way of stadium cameras which are already streaming visible information. Sports activities organizations can apply their experience to the visible information, and up to now there have been attention-grabbing outcomes. For instance, one firm extracted the angular velocity of a quarterback’s arm throughout a throw and confirmed that it is a sturdy predictor of elite efficiency.
Instruments like FitWise AI can change how groups perceive efficiency and harm threat. It’s a website the place the expertise is genuinely forward of the trade’s willingness to soak up it, and Kogan is just not ready for the trade to catch up by itself.
Discover these concepts and extra on our podcast
These are among the concepts we talk about on the JetBrains Analysis Podcast. If you happen to discover a query you may’t cease enthusiastic about, that’s by design.
All episodes can be found on YouTube and Spotify (and every other main podcast platform). We publish new episodes about as soon as a month.
