Summary

Gemini Deep Research targets the “zero to 50” work that otherwise consumes a weekend and 50–60 browser tabs. Running on a post-trained Gemini 1.5 Pro, it spends roughly 5 minutes turning a multifaceted question into a sourced report; ordinary Search remains better when users already know exactly what they want.
The editable research plan is both a steering surface and a contract for expensive agent work. Rather than interrogating a user who asks “tell me about batteries,” Gemini exposes its proposed facets and lets the user revise them conversationally—an “editable chain of thought,” as swyx put it, and Aarush agreed. Most users still press Start, but the plan explains why they receive the report they do.
The core technical advantage is iterative planning, not merely searching more pages. The model explores plan branches in parallel, reads results, notices gaps or contradictions, and chooses what to investigate next—for example, checking FDA policy after finding an EU ban. It then outlines, drafts, self-critiques, and revises, aiming to move beyond “high-level bullet points” into grounded second-order conclusions.
Long context preserves the active research journey, while RAG becomes overflow and long-term memory. Recent sources stay directly in context because users may ask fine-grained comparisons; material from “10 turns ago” can move into retrieval. Sridhar’s warning is that cosine-similarity retrieval weakens when a query contains multiple attributes, while newer long-context models remain effective even as their windows fill.
Latency is becoming a signal of effort, creating an eventual quality-versus-theater problem for research agents. Google tested a 15-minute “hardcore mode” but shipped roughly 5 minutes with an intended ceiling below 10; unexpectedly, users did not simply demand instant answers and could value seeing the work. The team has not found the perfect compute allocation between broader exploration and deeper verification, and users would probably press “max power” if offered.
Evaluation remains stubbornly human because the valid output space is too large for one benchmark score. Automated checks monitor plan length, iterative steps, and behavioral distribution shifts, while reviewers judge comprehensiveness, completeness, and groundedness across an ontology running from broad option discovery to narrow, deep investigation. “If I do great on HLE, that doesn’t really mean I’m a great deep researcher.”
The larger opportunity is a personalized, multimodal research agent connected to proprietary information, not just a better web summarizer. The team wants outputs adapted for a 15-year-old versus a PhD, generated as charts, maps, images, or interactive interfaces, and enriched by private documents and subscriptions. swyx’s investor-oriented call was unusually direct: this may be “the first real product-market-fit agent,” already plausibly worth $200 a month to some users and perhaps $2,000 when substantially better.

Deep dive

1. Deep Research begins by negotiating the work

Selvan’s product framing: Deep Research is a personal research assistant for going “from zero to 50 really fast” on a new topic. It browses for about 5 minutes, then returns a report that users can review, question, or reshape.
The boundary with Google Search is intent clarity. Search remains the destination when someone knows exactly what they need; Deep Research addresses multifaceted journeys that produce “50–60 tabs,” consume a weekend, and often end with the user giving up.
Before spending 5–10 minutes and meaningful compute, Gemini proposes a research plan. “Tell me about batteries” could mean innovations, chemistry, or a particular technology, so the model offers its first decomposition instead of forcing users through a sequence of follow-up questions.
swyx called the interface an “editable chain of thought,” and Aarush agreed. Early tests showed almost nobody editing, prompting the team to add an explicit button; even when users press Start like “I’m Feeling Lucky,” the plan remains a transparency mechanism and accepted contract.

2. Iterative planning turns web browsing into research

Sridhar’s technical account: the accepted plan contains parallelizable branches, which the model explores using two primary capabilities—searching and going deeper into a selected page. Crucially, it reads earlier results before choosing the next action.
The food-regulation demo carried the mechanism: if one search finds that the European Commission bans certain additives, Gemini can decide to check whether the FDA does likewise. Without that iterative grounding, Sridhar argued, the report stays incomplete and collapses into “high-level bullet points.”
Initial exploration is typically breadth-first, though the team does not hard-code that behavior. Gemini samples every plan facet, then double-clicks where a result is partial or sources conflict; once research ends, it constructs an outline, drafts the report, self-critiques, and revises.
The resulting milk-and-meat report did more than list rules: it inferred a philosophical split between an EU precautionary approach, which prefers to ban something despite inconclusive evidence, and a US reactive approach that allows it until harm is demonstrated. Selvan highlighted that as the desired “second-order insight.”

3. The report is a workspace, backed by tiered memory

Selvan divided follow-ups into three behaviors: recover a fact already encountered, launch another research pass for materially new scope, or directly edit the artifact by condensing, deleting, or adding sections. The side-by-side report and chat are designed to support all three.
If a user extends the comparison from the US and EU to Asia, Gemini decides whether existing research suffices or fresh browsing is warranted. All recently read sites remain available, so a missing detail can produce a quick answer without repeating the original 5-minute job.
When repeated research tasks threaten the 1 million–2 million-token context window, the team uses in-house retrieval. Its rule of thumb is to keep recent work in context for complex comparisons, then place material from perhaps “10 turns ago” behind RAG; related cohesive projects can stay in one thread because earlier niche discoveries may guide later searches.
Sridhar’s caveat on RAG: dot-product or cosine-distance retrieval struggles when the query itself contains several attributes. Newer model generations also preserve fine-grained recall deeper into long contexts, shifting the point where retrieval becomes preferable.

4. Web representation and multimodality remain practical trade-offs

Deep Research has both Markdown and HTML representations. Selvan said Markdown helps reduce the noise in pure HTML; the exchange cited JavaScript and Tailwind CSS as examples of that noise, while embedded HTML snippets may still need native handling.
Vision was not yet part of the product described. Selvan acknowledged the classic failure case—critical information trapped inside a JPEG—but argued that rendering pages adds latency for value concentrated in “a small part of the tail,” not today’s head use cases.
Fanelli’s pushback was that an agent already has a minutes-long latency budget. Selvan said models’ VQA capabilities are improving, but left rendering as a future trade-off rather than a current capability.

5. Evaluation starts with research behavior, not verticals

Output entropy makes evaluation “a hard one.” Automated raters can detect behavioral drift—changes in research-plan length, number of steps, planning time, or iterative-search depth—but those distributions only indicate that something changed, “for better or worse.”
Human reviewers therefore remain central, grading product-defined qualities such as comprehensiveness, completeness, and groundedness.
Sridhar’s evaluation ontology avoids vertical labels such as travel or shopping. One extreme is broad, shallow option exploration—finding many summer camps and summarizing each—while the other is narrow, deep understanding; comparison tasks and intermediate breadth/depth combinations fill the middle.
Compound projects test several patterns at once. Planning a Lisbon wedding may require researching planners, venues, and catering across ten subtasks. There is no hard conversation-turn limit, but most users currently do not go very deep. swyx argued that the completed document visually signals an endpoint rather than “the starting point”; Mukund agreed the UX could do more to invite continuation.

6. Useful latency is overturning Google’s speed orthodoxy

swyx identified a “perverse incentive”: agents that search 70 sites or run for an hour may look more capable even when 30 sources are irrelevant. Inefficiency can temporarily read as diligence, but he expects the honeymoon to end when users begin asking why the same quality was not delivered faster.
Google initially built two versions: a roughly 15-minute “hardcore mode” and the approximately 5-minute product it shipped. Selvan had asked engineering for a hard stop below 10 minutes, assuming users would abandon anything slower.
The surprise came when Jason Calacanis asked whether Google produced the answer in 10 seconds and merely delayed its display. That inverted the team’s experience with Assistant and other Google products, where lower latency reliably improved satisfaction and retention: here, visible effort itself carried value.
Selvan framed the compute decision as exploration versus verification. A question about Fed rate changes and middle-class household income should trigger precision and historical sourcing; birthday restaurants permit more leeway. The ideal agent infers that trade-off, because an explicit “max power” control would invite users to maximize everything.

7. Steering, specialization, and durable execution define the product moat

swyx’s strongest UX criticism was that users should be able to revise the plan while research runs. Devin exposes a live plan and accepts corrections mid-job; if research eventually lasts an hour, he argued, Gemini should behave like an intern who returns with findings, reports hiccups, and requests direction.
Selvan agreed that longer tasks would make in-flight steering more valuable: today’s few-minute research phase leaves little opportunity to intervene, but an hour-long task could let an agent return with findings and problems for direction. swyx’s proposed design was a live plan that updates and schedules the next job without locking chat.
Underneath the experience is an asynchronous platform: users can leave, close a computer, and receive a phone notification after completion. Five- or six-minute jobs inevitably fail, so the system preserves state, retries selectively, and avoids discarding completed research. Mukund said it is stable for hundreds of LLM calls and flexible enough for future hour- or multi-day workloads.
Shopping illustrates where specialization matters. Deep Research is weak for visually choosing shoes but compelling for HVAC systems, where specifications, voltage ratings, and finding contractors who can install them matter more than appearance. Selvan generalized this as “options exploration,” applicable equally to products, scholarships, or summer camps.

8. Better reasoning must lead to sourced novelty, not benchmark theater

The product is not purely Gemini 1.5 Pro: Selvan described it as a post-trained version and said there was no special access per se. He suggested the broad setup could be reproduced with tool calls and fine-tuning, including Gemma, while consistent planning and reliability require substantial post-training.
Thinking models introduce a second kind of inference-time compute alongside iterative web search. Mukund said they can draw more from model memory and produce stronger second-order insights, but sourced facts still need verification; even a correct remembered Fed statistic remains suspect without a .gov source. The challenge is balancing model memory with grounding. On moving beyond 1.5, his answer was “stay tuned.”
Generalizable iterative planning was the hardest modeling problem. Training separate traces for every domain or research ontology would be “nightmarish,” so the team emphasized data-efficient use of the model’s memory, data augmentation, and tuning post-training just enough to teach the behavior without erasing capabilities inherited from pre-training.
Benchmarks still help rally researchers—Selvan recalled MLPerf competitions rapidly improving TPU performance—but product validity is different. The team wanted to avoid optimizing for unnatural trivia such as “the day Kobe Bryant entered the league, who was the president’s nephew,” and Sridhar warned that text-output entropy makes verification and level-playing-field comparisons difficult.
Discovery beyond web synthesis requires both second-order reasoning and an environment that can test hypotheses. Code and math have sandboxes and verifiers; chemistry lacks an equivalent synthetic laboratory. Sridhar’s condition was explicit: an agent needs a playground, accurate feedback, and repeated experimentation before “new ideas” become more than unverified drift.
Selvan’s roadmap moves toward personalization and generative UI: a 15-year-old and a post-doc should receive different research reports, while charts, maps, images, and interactive structures should replace universal text documents.
The open web will eventually be the limiting corpus. Valuable industrial research lives in company documents, paid subscriptions, and private collections; yet Sridhar cautioned that it is still early to platformize agents into horizontal components. His builder advice was to choose one mission and “do that one thing really well.”

ImageNet Moment for Reinforcement Learning? [Prof. Jakob Foerster]Ravi Gupta - AI or Die - [Invest Like the Best, EP.411]