AscendC::Add(zLocal, xLocal, yLocal, TILE_LENGTH) — this single line is the entire point of the kernel. It runs on the Vector unit and adds all TILE_LENGTH numbers of xLocal and yLocal in one SIMD instruction, writing to zLocal. This is the Ascend equivalent of the GPU's out[idx] = a[idx] + b[idx] — except the GPU does one element per thread, while here one instruction does the whole tile.
very important
This is the whole "program" that runs on the AI Core — three lines
important
That's it. The entire "black box" is: chop the data into tiles, and for each tile, copy in → add → copy out.
important
So the chain, top to bottom: your PyTorch code → torch_npu → CANN → AI Core silicon.
Is the function wrapper_Custom_Tensor_add live within the tensor_npu library?
At some point, we have to actually access the data. PyTorch offers quite a few options for doing this.
We generally think of a kernel in PyTorch consisting of the following parts:
It’s a bit twisty and turny, so once you have some basic orientation about what’s going on, I recommend just jumping straight to the kernels.
end 3. The torch.add walkthrough
When you call a function like torch.add, what actually happens? If you remember the discussion we had about dispatching, you already have the basic picture in your head:
There may be multiple tensors which share the same storage. Storage defines the dtype and physical size of the tensor, while each tensor records the sizes, strides and offset, defining the logical interpretation of the physical memory.
end 2. Strides and Tensor/Storage
There’s also a little piece of metadata you might be less familiar with: the stride.
This is probably the most important mental picture to have in your head, if you’re trying to understand the way operators in PyTorch are invoked. We’ll return to this picture when it’s time to look more at code.
end 1. The dispatch picture
At the very most abstract level, when you call torch.mm, two dispatches happen:
Representations in the J-space can be used flexibly for many tasks—for example, once “France” has lit up in Claude’s J-space, the model can recall its capital, or its national currency, or the continent it belongs to.
a graph of connected nodes, each one is a concept
DenseNet itself is less often the default choice for general-purpose image classification or as a universal feature extractor in new large-scale systems, mainly because its concatenation-based connectivity can be memory-heavy at high resolution and scale.updf+1
important
Where the bottleneck actually is
very important
The converse is also interesting
important
This explains why contraction is the right inductive bias for reasoning
very important
Think about what happens when you solve a math problem step by step. Each reasoning step constrains the space of possible answers. You start with uncertainty (many possible values of x), and each logical step narrows it down until only one answer remains.
important
RL is fighting against the architecture
important
A huge fraction of what RL learns is "how to encode and decode continuous reasoning state into discrete tokens efficiently," not "how to reason."
important
And critically: you proved this preserves universal approximation. The constraint doesn't kill expressivity — it redirects it.
important
The match between problem structure and architecture structure would give you the same generalization benefit
super important
If this pattern holds, then current work on loop training pipelines might be in the "pre-ResNet" phase. The field is solving the hard problem of how to make destructive surgery work, when the real breakthrough will be discovering a non-destructive surgery that doesn't need the hard solutions.
very important
The reason ResNet worked at 100+ layers when previous architectures failed at 20 layers is almost entirely the additive structure. Everything else — batch norm, careful initialization — was supporting machinery. The core insight was: make the default computation be the identity, and let the network learn deviations from identity.
important
The retrofit playbook is incomplete without surgery recipes. Compare to the MoE retrofit story: Mixtral didn't just succeed because the training for upcycled MoE was figured out. It succeeded because the surgery — how to initialize expert weights from a dense base, how to insert routing layers, how to handle the load balancing transition — was also figured out. Those surgery techniques were as important as the training techniques, and they took multiple papers and years to mature.
important
So DeltaFormer is recovering some of the inductive bias of RNNs (state-tracking via recurrent memory updates) without abandoning the transformer's parallelizability — because the delta-rule recurrence can still be computed efficiently in a chunked/parallel form.
important
shorten CoT chains while preserving (or improving) reasoning quality, by moving the parts of reasoning that don't benefit from verbalization into latent space.
important
If looping-retrofit really does have LoRA-style adoption potential, then the research path that maximizes impact isn't "build the best looped model" — it's "build the cleanest retrofit recipe."
important
Retrofit-for-looping needs a specific pitch. "Better reasoning on multi-hop tasks at 5% of pretraining cost" is concrete and testable. "Generally better models" is not. The community adopts capabilities tied to specific problems they have.
important
Retrofit-for-looping needs an equivalent: a clean prescription that says "take base model X, apply procedure Y on data Z, train for N tokens, you get a looped version with capability gain Q." Retrofitted Recurrence is the closest thing today, but the recipe is still rough. The research opportunity is to refine it.
very important
If d is too small, beyond some loop count you're overwriting useful information faster than you're adding it.
very important
The key insight: answer-level diagnostics are cheap but coarse. They help the reflection LM form hypotheses but rarely confirm them.
Important
If you ask about coding now, she doesn't judge you; she gives you "inside info" to help you pass because she wants her candidate to win.
Important
Your updated, final belief after considering the evidence.
Important
and, this is why it is tempting for us, scientists, to think that this is the right and perhaps only way to tackle challenging problems. first, figure out precise mechanism by which phenomena behind these challenging problems happen, and then based on the uncovered mechanism, come up with solutions to these problems. surely once we know why and who problems manifest themselves, we can fix the problems.
Very very important argument. This is similar to the intuition of Ilya about Deep Learning. In the past, the reason why researchers prefered classical ML is because they can theoretically prove understand and prove their mechanism. But Ilya think Deep Learning is perfect because we cannot model them.
The Limitation: Its greatest weakness is its inability to perform extrapolation.
Important
It is fundamentally an exploratory approach.
Important
It is fundamentally an exploitative approach, seeking the "best" answer according to the flawed map it's given.
Important
The system's potential is demonstrated through concrete validation in three biomedical areas: drug repurposing, novel target discovery, and explaining mechanisms of antimicrobial resistance. For instance, it proposed drug candidates for acute myeloid leukemia that showed tumor inhibition in vitro, showcasing its ability to produce genuinely valuable and original scientific insights. This work frames a new vision for AI: a system that can navigate the high-dimensional space of existing scientific knowledge to discover the unknown.
align
Foundational Models and Scientific Discovery: The AI Co-Scientist
align
Proposes a multi-agent system based on Gemini 2.0 to automate hypothesis generation in scientific discovery, using a "generate, debate, and evolve" approach.
align
A crucial operational maxim is to "be stubborn on the vision and flexible on the details," acknowledging that this flexibility is necessary because the world is changing [1].
Important
change very slowly
Important
points of stability
Important
safety mechanism — it builds understanding rather than surface mimicry.
Important
To understand, we need to find an internal model that can generate those patterns.
Important
It becomes more honest as it becomes more powerful.
Important
Formally, it minimizes something like log loss or cross-entropy between predicted probabilities and actual observed outcomes — a metric that rewards calibrated truthfulness.
elaborate
In standard RL or reward-maximizing setups, an agent can over-optimize the reward proxy (Goodhart’s law) and exploit loopholes. A Bayesian, inference-only system doesn’t optimize for outcomes—it merely models them. The ensemble of hypotheses provides natural regularization against runaway single-metric optimization.
elaborate
Toronto team’s output is on general AI capabilities (LLMs, generative models, agent learning), rather than domain-specific science tools.
Important
the prevailing corporate governance models are fundamentally ill-suited to the unique, long-term safety requirements of AGI development.
Important
Institutionalizing the Vision: The Role of LawZero
Important
The "Guardrail" Function: Using Safe AI to Monitor Unsafe AI
Important
non-agentic. Instead of being trained to take actions to achieve goals, it is designed to explain the world from observations.
Important
graph Laplacian
graph Laplacian of the causality graph for natural language, which is a tree, is zero?
boundary value problem on a graph
Important
discrete Poisson equation
need elaborate
Adapting Geometry for Directedness
I have a feeling that we can learn the "shape" of the DAG if we investigate it under the lens of Discrete Differential Geometry
The GFlowNet "flow matching" loss condition can be interpreted using this language, relating the flow into a state to the flow out of it.
elaborate
dynamics
Important
The production rules of a grammar are the constraints that force the data to a low-dimensional manifold. A molecule's grammar, for instance, prevents the vast majority of random atom-and-bond combinations from ever being formed, concentrating all valid molecules onto a specific, structured surface.
makes sense
improving predictive models of the world
elaborate
Where MuZero actually sits
seems important
diffusion needs extra machinery to cope with discreteness
need elaboration
Native fit for combinatorial/structured objects.
need elaboration
Discrete/combinatorial data is awkward.
??
these biases might over-constrain search and prevent discovery of novel molecules
Super important
Effect: credit from terminal rewards can reach all ancestors in one step through the balancing equations — no need to wait for step-by-step temporal backups.
this is so important
So instead of waiting for a single reward at the end of a trajectory, each local edge is trained to satisfy a conservation law consistent with terminal rewards.
Important
Aggregating Multiple Trajectories
need elaboration
Local Signal
need elaboration
RL would need repeated training or MCMC-like sampling for each new query.
elaborate
Efficient Exploration of Combinatorial Spaces
??
GFlowNets use flow-matching (inflow = outflow) as a local conservation law — mathematically simpler, more stable, better credit assignment.
elaborate
Produces a distribution over solutions — crucial for scientific discovery, Bayesian inference, or multi-modal problems.
elaborate
Inefficient Exploration
elaborate
Single Trajectory Credit Assignment
elaborate
interpolates/extrapolates reward structure: If many molecules with a substructure A had high reward, The policy will bias flow toward other molecules that also contain substructure A — even if it never explicitly saw them.
important
Generalization across modes
elaborate
Every new sample requires running a potentially long Markov chain. No benefit is carried over between runs.
elaborate
The above two equations are forced to be consistent (i.e. there is an FFF that gives rise to both PBP_BPB and PFP_FPF) when FFF satisfies the flow-matching constraint (the amount of entering flow equals the amount of outgoing flow), which is not necessarily true when FFF is estimated by a neural network that is being trained (and we have not fully completed training and brought the training loss to 0 everywhere).
can we force this constraint by design?
System 1 (policy net) learns to approximate those judgments instantly.
so planning ahead skill emerge?
They don’t use explicit grammar rules. Instead, the grammar is implicitly encoded in the model’s weights.
important
The world is combinatorially large (endless possible combinations). Humans generalize well by reusing parts without needing to see every possible combo in training.
Is this what happens to LLM and possibly stable diffusion?
AlphaZero → trained on massive self-play, now plays “intuitively” via forward neural evaluation (System 1).
need elaboration
A small model might need symbolic rules to parse and generate. A large LLM trained on trillions of tokens can produce fluent sentences directly, implicitly encoding the grammar.
important. need elaboration
Represented symbolically, compositionally, or in step-by-step logic. Conscious and accessible: you can bring it into working memory, verbalize it, and explain it.
elaborate
Drug discovery models (like GFlowNets, graph diffusion, VAEs) exploit this structure: they learn to navigate and generate only chemically valid molecules rather than arbitrary graphs.
important
Sequential Construction
any relation to the sequentially adding noise step-by-step in diffusion model?
2. Initially introduced for active learning in discrete spaces
need elaboration
Instead of treating a design as one huge blob (like a voxel grid), CGID represents objects as a composition of functional parts.
similar to the compositionality spirit of GflowNet
1. Does a GFlowNet need training data points? Not necessarily. If you have a reward function that you can compute for any candidate object, the GFlowNet can train purely by sampling objects (even ones it has never seen before) and scoring them. This is why they’re attractive for drug/material design: you don’t need a huge dataset of known good molecules, just a scoring function (simulator, energy model, ML predictor) to evaluate new candidates.
important
Data-efficiency problem: to represent rare but valuable cases, you need a lot of data containing those cases.
important
By itself: can sample realistic shapes, but functionality is random (not aligned with tasks).
??
Rather than finetuning the whole diffusion model, it optimizes the conditioning embeddings (e.g., CLIP embeddings) with feedback from simulation performance.
??
soft robots
what is
LLM priors keep evolving
maybe Claude playing Pokemon indeed does this?
But if they have an energy landscape (a surface with valleys and peaks corresponding to good vs bad configurations), then you don’t need to brute force. You just need to learn the gradient — the direction downhill toward stable/valid solutions.
interesting. Need elaboration
This assumption is empirical: generative models (autoencoders, GANs, diffusion models, LLMs) succeed precisely because such structure exists and is learnable.
evidence that generative models can capture these structure just by observation, without interacting with the physical world?
In mathematical terms, if the total search space is size NNN, but the effective dimensionality of “plausible” or “stable” solutions is only size M≪NM \ll NM≪N, then we say the domain has exploitable structure.
manifold
stable, reusable patterns (games with coherent strategy, protein physics shaped by evolution).
elaborate
What’s strong about P
it makes me think about the general & universal PS system with core Transformer and different adaptors for each task. Can we build a similar universal planning system that shares the same core LLM?
i'm not sure if this is blackpill or whitepill, but my there are a heap of new papers along with my own experiences that are showing "best of N is all you need" for most problems as long as: - sufficient core knowledge was included in the training data - the model is sufficiently large / you use more than 1 model to promote reasonable idea diversity - N is sufficiently large for the complexity of the problem at hand - you have some reasonable discrimination process at the end to determine / approximate "best" result we really haven't come close to leveraging the full potential of existing models, and the antiquated sampling process / approaches are the single biggest culprit
what if we include some sampling distribution on the output side?
This is already happening with smaller labs: open-source models are trained partly on outputs from frontier APIs.
genius
What Culture Does for Humans
is culture an emerging capability under the survival pressure?
Aidan hints at this: culture is to humans as wrappers are to models
what does this mean?
If they achieve network effects (large user base, ecosystem of integrations, proprietary data flows), they could become the “operating system” for interacting with AI.
important
Consumers
can researcher be customer?
5. Cultural and Human Analogies
elaborate
Both groups are in a race to the bottom of commoditization, where only brand, UX, or network effects (e.g., Perplexity’s model-switching) can provide some edge
why?
Wrappers that add UX, workflow, and integration value may persist, but rarely become billion-dollar companies.
why?
Even if scaling is more generally efficient than search, search allows for quicker intelligence in narrow domains. Training larger foundation models is slow. With search, you don’t have to wait.
this is the key lesson
I like research topics that are simple, general, and stand the test of time, and I try to avoid projects that are complicated, task-specific, or short-lived.
elaborate I guess general research topics are more fundamental and tend to have longer lifespan?
Most people (including me) would benefit greatly by spending more time on idea selection, since doing this well is a huge multiplier on research impact. Conversely, working on a narrow topic with little headroom caps the impact of the project, no matter how well it is executed.
elaborate
A good suggestion from a friend is to either (1) work on a hot topic and do it better than everyone else, or (2) work on something that might become the next hot topic. Strategy 1 is lower risk and requires working very hard. Strategy 2 is higher risk but has potentially very high reward.
elaborate
🔹 General Research Topic → Broad Benchmark Suite
this is what happen to the PS project?
They scale with the frontier: as foundation models improve, broad topics grow in importance, while narrow ones fade.
why these scale with frontier?
It is an open question why exactly scaling works, but here are two hand-wavy reasons. One is that small language models can’t memorize as much knowledge in their parameters, whereas large language models can memorize a huge amount of factual information about the world. A second guess is that while small language models are capacity-constrained, they might only learn first-order correlations in data. Large language models on the other hand, can learn complex heuristics in data.
important
Intuition 3. Tokens can have very different information density, so give language models time to think.
important I guess he realized this by examing the behaviors and output of LLMs. Then what happens next is that it inspires him about the intermediate thinking chains of thoughts
The solution to this is to give language models more compute by allowing them to perform natural language reasoning before giving the final answer.
nice
You can imagine that if you’re ChatGPT, and as soon as you have to see the prompt you have to immediately start typing, it would be pretty hard to get that question right.
important
Intuition 2. Learning input-output relationships can be cast as next-word prediction. This is known as in-context learning.
need elaboration
This is an interesting example of how a simple objective, when combined with complex data can lead to highly intelligent behavior (assuming you agree that language models are intelligent).
nice observation
It is important to understand that this does not mean LLMs will be gods producing 100x code, because virtually no domain that software engineering is useful has a perfect oracle. A perfect oracle is a type of feedback where you are given a “correct/incorrect” answer every single time, and they almost only appear in games as real world typically doesn’t have perfect models of correctness. Winning or losing a game is a perfect oracle, as well as creating a program that can pass the judge in a competitive programming contest.
important and impressive advice
It is the limit that tells us what we cannot implement via LLMs, and it cannot be solved with agentic approaches.
important
However, in scientific innovation, we are in a totally different realm where we only care about solving a single problem (train=test!) because it’s an unsolved problem and potentially extremely valuable.
.
overfitting
.
any solvable problem that fits those five properties will be solved in the next few years
important
Speed of iteration
.
If you consider the history of deep learning, we have seen that virtually anything that can be measured can be optimized
great
If you consider the history of deep learning, we have seen that virtually anything that can be measured can be optimized
great
Prefer methods that pass a scaling test: their delta is flat or increasing as you go from S→M→L models.
maybe we can just plug in models of various scales and see how the performance project?
Extra compute/search
I guess this is needed when generator-verifier gap exists?
Hard constraints:
hard domain-specific constraints
If a method is just a shortcut for weak models (e.g., brittle augmentations, dataset-specific features), it will fade.
important
If a method is just a shortcut for weak models (e.g., brittle augmentations, dataset-specific features), it will fade.
important
Method: anything you wrap around or plug into the core model: data curation, training tricks, inference procedures, retrieval, tools, constraints, rewards, routing, etc.
maybe each AI startup out there is a different wrapper, specific for each domain?
Why methods often fade away
maybe because these methods don't take scaling into account?
🔹 What does “method” mean here?
So there are 2 types of hammers: - architectures that scale well with compute & data - methods that scale well with compute & data It might be interesting to bring these compute-scalable methods to integrate in architectures of other domains? What are historic examples for these compute-scalable methods? Are they general-purpose?
4. Broader principle
best
Build Small Internal Tools
need elaboration
summaries of what worked, not the nuanced landscape of the problem
important
you’re not competing — you’re complementing.
important
adjacent signals: lots of startups are already building agent frameworks (LangChain, AutoGPT, CrewAI). Their bottlenecks hint at where research could contribute
great advice
2020: you can cast any language task as sequence prediction and learn it via pretrain + finetune 2021: scaling to GPT-3 size enables doing arbitrary tasks specified via instructions 2022: scaling to GPT-3.5/PaLM size unlocks reasoning via chain of thought 2023: LLMs themselves can be a product a lot of people will use 2024: to push capabilities past GPT-4, scale test-time compute
there is a dominant trend and scaffolding structure here
Old habit: Just grab a standard benchmark (e.g., GLUE, ImageNet, MMLU) and test your method there. Problem today: Those benchmarks might not stress the thing your method is designed for. You’ll conclude your idea “doesn’t work,” when in fact you just used the wrong test.
important
Now: Large language models are so capable and multi-task that whether a method works depends a lot on which dataset you test it on.
elaborate on this
AlphaEvolve
what is it?
The field rewards researchers who can translate expensive experimentation into deep, portable ideas.
nice
AcknowledgementsThis blog post features contributions from Gabriel Ilharco. I would like to thank Hattie Zhou, Nelson Liu, Noah Smith, Gabriel Ilharco, Mitchell Wortsman, Luke Zettlemoyer, Aditya Kusupati, Jungo Kasai, and Ofir Press for their valuable feedback on drafts of this blog post.
this guy certainly not feel shy when asking people for feedback on his work, no matter it's compicated as doing research or as simple as writing a blog
It is important to note that there is neither right or wrong nor good or bad research style.
important
He builds hacks, understands the deep relationships of how his hack affects the system, and then extracts this insight in the most minimalistic and well-formulated way possible along with his practical hack.
This is so good advice
Navigating this uncertainty is best done through fast iterations and balancing multiple projects to maximize the chances of a big success.
important
The learning from one doesn’t always transfer, because the ideas are not anchored to a common goal.
very important
✅ Correct in spirit: Solving “make X work for the first time” usually looks like jumping from essentially no working method to a viable solution (say 0% → 70%).
this looks far more impressive than incremental works
X = real-time object detection on a single GPU.
nice example
If they join a crowded subfield late, the low-hanging fruit is gone.
I read somewhere that great researchers never stay in a crowded place. They move on to the "next big thing". Related posts on low-hanging fruit: https://ai.engin.umich.edu/2023/08/17/eight-lessons-learned-in-two-years-of-ph-d/ https://www.notion.so/From-Michael-Nielsen-Principles-of-effective-research-25199f12ef0d8040be1ceb0d8b16cc67?source=copy_link#25199f12ef0d80649850edcef18f5519
Pick a meaningful dimension of stress (tokens, latency, noise, compositional depth, generalization to new domains, etc.).
how to select out a "meaningful dimension" is an important follow-up question by it own. I have several points to add: - the dimension should also involve "feasibility" that fits our research condition - maybe to find out those "meaningful dimensions", we can look very top-down from applications. We can ask, if this dimension is extended, which kind of application can benefit. And with this, we quickly realize that "context length" is a super influential dimension, benefiting various applications. - how about borrowing from related domains like Jinwoo did in TokenGT?
Relevance to the field’s trajectory The capability should connect to active conversations in the community. Example: In 2020–2022, instruction-following was “in the air” because GPT-3 showed emergent abilities, but not controllability. So InstructGPT’s capability (follow human instructions) was both new and natural.
I guess ChatGPT can help me with identifying the next natural new AI capabilities to work on
Do you want me to break down how to tell, when reading a new paper, whether it’s a “first-time X” paper or a “make X better” paper? That skill will help you classify work quickly.
so there are many dimensions for "working better"? e.g more stable, more scalable, ...
Alternatively, armed with the knowledge gained from working on the first idea, you might move on to a different idea aligned with the same goal, with a higher chance of success.
so knowledge from working on the first idea leads to higher chance of success in the second idea? So I guess it's all about fail fast and then iterate?
If you are thinking in terms of ideas, you’d be easily frustrated and might give the idea a few more attempts before finally giving up and moving on to another, possibly unrelated idea, repeating the same process
"unrelated idea" is an important point
John Schulman argues that goals have more longevity than ideas
did he?
The main issue is that ideas have a very short lifespan; an idea is unlikely to work at first, might not be novel enough, might be easily scooped
elaborate
When you’re starting in a new area without a full understanding of the challenges or limitations, It is very tempting to run after a sole idea that you think will work.
why?
What Yang was talking about is reading enough papers to cover most of the literature in your area. Needless to say, it is not only about the paper count you read, although the paper count can serve as a good indicator of how well you are engaged with the literature in your research area.
the final goal is to cover understanding of your literature
enabling you to make larger leaps of progress.
why?
new AI capabilities
this is too broad. Is there any constraints to narrow down the set of new AI capabilities that we should vision in the scope 2 years?
showing how to do X
so the problem of "making X works for the first time" is already resolved?
choosing a different problem from the rest of the community can lead you to explore different ideas.
how different here exactly?
initial exploration
what is this?
Goals also make it possible for a team of researchers to work together and attack different aspects of the problem, whereas idea-driven research is most effectively carried out by “teams” of 1-2 people.
why??
On the other hand, with goal-driven research, your goal will give you a perspective that’s differentiated from the rest of the community. It will lead you to ask questions that no one else is asking,
why??
To make breakthroughs with idea-driven research, you need to develop an exceptionally deep understanding of your subject, and a perspective that diverges from the rest of the community—some can do it, but it’s difficult.
why??
you test a variety of existing methods from the literature, and then you develop your own methods that improve on them
(1) Which existing methods to test? (2) Why we have to test them before we build our own methods? What does it mean by "improve on them"?
I’ll take goal-driven research to mean that your goal is more specific than your whole subfield’s goal, and it’s more like make X work for the first time than make X work better.
what does this mean? "your goal is more specific than your whole subfield's goal" how does that align with "make X work for the first time than make X work better"?
solve problems that bring you closer to that goal.
how to identify problems that if we solve, could bring us closer to that goal?
Follow some sector of the literature.
what does this mean exactly?
This AI-driven auto-scheduling is a massive time-saver, eliminating the tedious manual process of trying to Tetris tasks into your day.
important
A 40 hour time-blocked work week, I estimate, produces the same amount of output as a 60+ hour work week pursued without structure.
bring structure into life always wins
Combine with MVO first → secure a skeleton first, then apply Pareto logic for depth.
important
5. Pareto 80/20 Why #5: Still powerful — often a few papers/experiments/figures give most of the insight. But research is exploratory, so it’s easy to misjudge which 20% matters most until later. Works best if combined with checkpoints. Example: In experiment runs → 2–3 baseline setups cover 80% of insight; no need to sweep every hyperparameter.
CS 197: Computer Science Research Step 1: Performing a literature search Keep track of how much you’re learning about the design axes as you consume additional papers. Typically, you’re learning the most at the very beginning, and the amount per paper starts going down after five papers or so.
6. Progressive Layering Idea: Build the output in layers: skeleton → basic fill → deeper detail → polish. How to apply: Do a quick pass that touches everything at a shallow level, then loop back to deepen. Good for: Ensuring balanced coverage and avoiding “holes.” Limits: Requires resisting the urge to perfect one section before moving on.
important
4. Greedy Value-per-Time Idea: Always pick the next piece of work that gives the highest value for the time spent. How to apply: For each possible action, estimate: “How much value will this add?” / “How long will it take?” → Do the one with the best ratio. Good for: Gradually building up value in the most efficient order. Limits: Estimation can be rough; might overlook long-term gains.
need elaboration
3. Time-Boxed Checkpoints Idea: Divide your time budget into checkpoints (e.g., 25%, 50%, 75% of T). How to apply: At each checkpoint, stop and take stock. Freeze the current delivery so it’s already usable, then decide whether to add or polish. Good for: Avoiding the trap of chasing “better” until the very last minute. Limits: Requires discipline to actually stop and review.
need elaboration
Lesson: the hidden blocker (“taxonomy dimension unclear”) surfaced cheaply at L0.
important
Without early reality checks: you drift → big risks discovered late (wasted days). With L0/L1 reality checks: you “fail fast, small, and cheap,” so every later step builds on a working scaffold.
fail fast
You fix it before scaling up.
important
utility proxies (does this draft/pipeline already work in some small way?)
super important
8) Common pitfalls (and the fix)
elaboration
You’ll increase 1–2 knobs per level, never all.
important
usability-focused.
elaborate on this
Converts one far-off, sparse reward into 5–6 immediate feedback signals.
important
Break your open-ended milestone into layers, each with a feedback proxy:
important
Solution: tie every proxy to usability: “Can this artifact unblock the next step?” If no → proxy doesn’t count.
need elaboration
Very sharp connection, Dat 👍 — you’re right: open-ended research tasks suffer from the same two problems as reinforcement learning (RL):
both RL and research are hard. I guess feedback is the bottleneck for both.
Get to “usable” fast → then refine only if it proves valuable.
gold
👉 Philosophy: The goal of early outputs isn’t to be complete — it’s to learn faster.
important
Philosophy: Closure is manufactured, not discovered. If you don’t impose an end, the task never ends.
gold
If you ship something usable by the deadline → ✅ pipeline works. If not → ❌ pipeline broke (e.g., stuck in collection or polish).
important