Small questions. Big disagreements.
Eight questions from the guide. Pick a side, compare notes with other readers, and change your mind when the evidence moves.
No account needed. Reader opinions, not a scientific survey or an official verdict. How voting works ↗
Results refresh every 30 seconds. One current answer per poll in this browser.
Will Transformer-like models lead most NLP benchmarks on January 1, 2027? Make a prediction before the clock runs out.
This is a reader forecast, not the wager’s official resolution. Our countdown uses midnight UTC as a stated convention; the original terms do not specify a time zone. Voting on this forecast closes at that countdown deadline.
Read the resolution policy →Which definition of “Transformer-like” would you use to settle the debate?
The guide has three explicit conventions. Including hybrids or pure linear attention can change which architectures qualify without changing a single benchmark score. These are choices about the category, not claims about a universal taxonomy.
Compare the three definitions →When you look at Jamba’s original architecture, which description feels most useful?
The original Jamba configuration interleaves one attention layer with seven Mamba layers. It is excluded by our narrow convention and included by the other two. This poll asks how you would describe it; the mechanism itself is documented.
Meet the original Jamba →You’re funding the next architecture experiment. Pick the branch you most want to see explored.
These are the four documented architecture branches in this guide. Mixture-of-experts is a separate design choice: it can coexist with attention or other sequence mechanisms. A research preference is not a leaderboard ranking.
Browse the architecture families →Imagine building a reading assistant for long documents. Which improvement would you prioritize?
Quality, speed and memory are different axes. Our historical long-context basket measures reported answer quality using F1; it does not score all of these priorities or isolate architecture from training and hardware.
Explore the long-context comparison →Our long-context example changes from 2 of 5 qualifying task leaders to 5 of 5 when hybrids count. What is your main takeaway?
In Table 3 of the 2024 Jamba paper, Jamba leads three tasks and Mixtral leads two. The switch comes from reclassifying Jamba’s wins. It is a bounded, historical comparison, not a result across the current field.
Change the rules yourself →Pick the piece of evidence you would find most persuasive about a challenger architecture.
Our guide separates documented mechanisms from benchmark outcomes. Reported scores have limits: training data, compute, prompting and task selection can all affect a comparison. Choose your strongest signal, even if you value several.
See how the guide handles evidence →If this architectural debate were a follow-up paper, what would you put on the cover?
A bit of levity after the footnotes. These are playful, invented titles, not research papers or evidence about model performance.
Read the real research titles →Votes are stored in Supabase. A random browser cookie remembers your answers for up to one year; no name or email is requested. You can remove your vote below each result. Counts represent ballots, not verified unique people, and percentages may not add to 100% because of rounding.