ML etc.¶
......
robotics_control_optimization vs CV -> LLM¶
https://www.zhihu.com/question/22031360/answer/20092442
------
https://www.zhihu.com/question/40120261/answer/3605671072
Given that dynamic control has so many advantages, why are most industrial robot products based on kinematic control?
Because it has higher implementation difficulty & greater cost & for most industrial applications kinematic control already meets the requirements, industrial robot systems generally adopt kinematic control — avoiding "using a cannon to kill a mosquito" — to save costs.
Kinematic control systems are relatively simple, easy to implement, debug, and maintain, and possess high stability and robustness, able to ensure the precise position control required in industrial tasks.
Kinematic control technology is already very mature in industry, with a well-developed software and hardware ecosystem that complies with industry standards and practices, further promoting its widespread adoption. Any product relies on its ecosystem — for example, the position of WeChat in China's communication software.
https://www.zhihu.com/question/530581913/answer/3607297565
[https://www.zhihu.com/question/633124010/answer/3598256957]
------
[https://www.zhihu.com/question/402251886/answer/1344947652]
Model-based RL is no different from traditional optimal control; it's just a different name. Although RL sounds impressive claiming it can produce a global control law, the vast majority of systems actually cannot reliably do so. Hence, in recent years, DRL has advocated using deep nets to fit such a control law (or the value function).
As for whether it works, that depends on everyone's "faith" in ML; as for MPC, it is much more down-to-earth — numerical optimization at least returns a local optimum, and sometimes it can be proven that a local optimum already achieves the desired control effect. By exploiting the sparsity of the optimization problem, MPC optimization algorithms can be made very real-time. Of course, for complex non-convex problems, relying on local numerical optimization alone is also insufficient.
------
For MPC-based vehicle controllers, given the powerful on-board chip compute nowadays, the computation of MPC is negligible. Implementing MPC attitude control on a drone, the flight controller was written on a Raspberry Pi; with a 30-step prediction, solving the QP takes less than 1ms, which is real-time enough for me.
In this era of explosive compute, we can freely embrace high-compute platforms and high-compute algorithms. Claims that MPC and RL are impractical are old-generation thinking. If I recall correctly, Boston Dynamics' quadruped and biped robots, and MIT's open-source robot dog, all run high-dimensional MPC in real time (?).
(DeepMind) Towards General and Autonomous Learning of Core Skills: A Case Study in Locomotion [https://www.semanticscholar.org/paper/Towards-General-and-Autonomous-Learning-of-Core-A-Hafner-Hertweck/afeffb9e05d89b2ac806282d3ed4366d67e4392e]
This DeepMind work uses no MPC at all; it simply designs a multi-level reward function and keeps training. DeepMind has always been committed to general intelligence (?).
By comparison, ETH's structure is far more complex, still requiring MPC and a bunch of traditional control methods. In terms of results, ETH seems a bit better (judging by terrain), but still appears no match for Boston Dynamics' hand-tuned MPC.
- https://leggedrobotics.github.io/rl-blindloco
- https://leggedrobotics.github.io/rl-perceptiveloco
- https://arxiv.org/abs/2309.14246v2 https://sites.google.com/leggedrobotics.com/risk-aware-locomotion Learning Risk-Aware Quadrupedal Locomotion using Distributional Reinforcement Learning
https://github.com/TinyMPC/TinyMPC
------
https://zhuanlan.zhihu.com/p/504256899
https://zhuanlan.zhihu.com/p/115726898
https://www.zhihu.com/question/650407644/answer/3448786787
degrees of freedom
https://www.zhihu.com/question/597238433/answer/3080541702
============
log¶
RWKV-7 as a meta-in-context learner The model's internal world must continuously fit the external world From this first principle (?) one can directly write out the exact formula for RWKV-7 https://zhuanlan.zhihu.com/p/9397296254
https://www.rwkv.com/images/RWKV-7.png https://github.com/BlinkDL/RWKV-LM/tree/main/RWKV-v7
- RNN with great LLM performance, which can also be directly trained like a GPT transformer (parallelizable). We are at RWKV-7 "Goose".
- combining the best of RNN and transformer - great performance, fast inference, fast training, saves VRAM, "infinite" ctxlen, and free text embedding. Moreover it's 100% attention-free, and a Linux Foundation AI project.
I believe RNN is a better candidate for fundamental models, because: (1) It's more friendly for ASICs (no kv cache). (2) It's more friendly for RL. (3) When we write, our brain is more similar to RNN. (4) The universe is like an RNN too (because of locality). Transformers are non-local models.
All the trained models will be open-source. Inference is very fast (only matrix-vector multiplications, no matrix-matrix multiplications) even on CPUs, so you can even run a LLM on your phone.
RWKV is parallelizable because the time-decay of each channel is data-independent (and trainable).
RWKV-7 uses parallelized mode to quickly generate the state, then use a finetuned full RNN (the layers of token n can use outputs of all layer of token n-1) for sequential generation.
Were RNNs All We Needed? https://news.ycombinator.com/item?id=41732853 https://arxiv.org/abs/2410.01201
Dec 18, 2024
============
Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning
The most interesting thing about this is the way it was trained using synthetic data, which is described in quite a bit of detail in the technical report: https://arxiv.org/abs/2412.08905
Dec 23, 2024
============
One major problem OpenAI faces is a lack of diverse, high-quality data; the public internet does not have enough data for training. Another problem is brain drain.
https://openai.com/12-days/?day=12 https://slashdot.org/story/24/12/22/0333225/openais-next-big-ai-effort-gpt-5-is-behind-schedule-and-crazy-expensive
Dec 31, 2024
============
Does this mean you don't need large GPU clusters for frontier LLMs? No but you have to ensure that you're not wasteful with what you have ... https://www.solidot.org/story?sid=80186 DeepSeek claims its new model cost only $5.5 million to train
https://github.com/orgs/deepseek-ai/repositories?type=all
V2 https://www.zhihu.com/question/655172528/answer/3490846800
Jan 6
====== ======
DeepSeek-V3 currently offers the best price-performance ratio https://github.com/deepseek-ai/DeepSeek-V3/blob/main/DeepSeek_V3.pdf
Math¶
-
- o3 scored 96.7% in the AIME math competition (up from 83.3% for o1), improving by 13.4 percentage points.
-
- o3 achieved 87.7% accuracy on the PhD-level GPQA Diamond benchmark, well over the average of 70% for PhD experts.
-
- For the Epic AI Frontier Math Benchmark, o3 is the only AI model to exceed 25% accuracy, while others typically score under 2%.
-
Extremely Expensive: The cost of o3 is currently 1000 times that of o1. According to the ARC-AGI testing standards, solving one problem with o3 in low mode costs around $20, but in high mode, the cost can skyrocket to around $3,440 per task. For instance, asking "Which is greater, 9.09 or 9.11" would set you back $1,600,250 for a total of 400 + 100 tasks in the ARC-AGI test.
https://x.com/asankhaya/status/1858714149557334481
Competitive coding¶
Competitive coding is usually difficult, but not complex. Software Engineering is usually complex, but not difficult. https://www.reddit.com/r/programming/comments/1hir7lb/the_new_openai_model_o3_scores_better_than_998_of/
SWE-bench¶
Can Language Models Resolve Real-world Github Issues? https://openreview.net/forum?id=VTF8yNQM66
https://www.reddit.com/r/LocalLLaMA/comments/1hc276t/gemini_20_flash_beating_claude_sonnet_35_on/
Jan 24
============
// 1. jump-game
https://sara-hy.github.io/2019/01/10/leetcode-JumpGame/
https://builtin.com/software-engineering-perspectives/jump-game-leetcode
// 2.2 Gomoku self-play https://ne7ermore.github.io/post/alpha-zero/
https://www.nature.com/articles/nature24270.epdf?author_access_token=REDACTED
// 3. https://techxplore.com/tags/board+games/
https://www.science.org/doi/10.1126/sciadv.adg3256
** https://deepmind.google/discover/blog/
I have already been eliminated ......
============
// - 1. https://openreview.net/forum?id=XYK1eGjahp
https://m.youtube.com/watch?v=o7uac6DuzcQ
How our work differs from the above-mentioned results: Many of the above papers are focused on problems in P or P/poly, while 3-SAT is an NP-complete problem. It is widely believed that P is a strict subset of NP, and it is not known whether NP is a subset of P/poly. In other words, our results are not comparable to these earlier results.
Meanwhile, Pérez et al. (2019), Li et al. (2024), and Merrill & Sabharwal (2024) also show that Transformers can simulate single-tape Turing Machines (TM) with CoT and can theoretically be extended to arbitrary decidable languages. However, these constructions require at least one CoT token for every step of TM execution.
By contrast, our theoretical construction demonstrates that, for certain classes of formal reasoning problems, Transformers can simulate algorithmic reasoning traces at an abstract level with drastically reduced number of CoT tokens compared to step-wise emulation of a single-tape TM. At each CoT Step, our construction performs deductive reasoning over the full input in parallel while any single-tape TM must process each input token sequentially.
Furthermore, the CoT produced by our theoretical construction abstractly represents the human reasoning process of trial and error, as demonstrated in Figure 1.
// - 2.
For example, a 2023 ICLR paper mentioned that the Transformer architecture is capable of implementing several different linear regression algorithms; this positive result again demonstrates its generality. https://x.com/xleaps/status/1627873094814531584
& https://twitter.com/xleaps/status/1627844991824297984
An "out-of-the-box" network example: when training a neural network to compute modular addition a+b = ? (mod c), the network suddenly achieved 100% accuracy at some point. Computing modular addition using the Fourier transform. This algorithm can be proven correct & is counter-intuitive to humans (?).
In this example, about 1400 steps of training moved the network from simply memorizing the training data to the Fourier transform; about 9000 steps of training made the Fourier transform perfectly implement modular addition, and the remaining few thousand steps cleared the network's prior memory, transferring all weights to the subnetwork that implements the Fourier-transform algorithm.
LLMs can be viewed as heuristics we adopt because we are unable to control emergence.
// - 3. SFT overfitting : short term RL generalize : long term https://pbs.twimg.com/media/Gje3D_AbYAAuBfP?format=jpg&name=large
============
https://blog.computationalcomplexity.org/2025/02/research-then-and-now.html
Low hanging fruit : computational complexity was just 20 years old when I started grad school ... the P v NP problem You didn't need to know deep math and most graduate students could follow nearly all of the 47 talks at my first conference ( STOC 1986 ), not likely in the STOC 2025 papers.
Now you need to spend time climbing the trees and going down deep branches to find new problems that only people on nearby branches would care about or even understand.
But who knows, AI may soon climb those branches for you.
============
// - 1. AI-induced cognitive decline: requires quantitative measurement https://www.solidot.org/story?sid=80529
// - 2. https://x.com/wwwyesterday/status/1886363047255773357
// - 3. DeepSeek-R1 : jailbreak 100% https://x.com/rohanpaul_ai/status/1886025249273339961
// - 4.
Some analysis suggests that the NVIDIA GPUs used by DeepSeek were about 10-30% cheaper on the market than the cutting-edge products used by US companies. DeepSeek used 2000-3000 H800 (the China-special edition of the H100) chips to develop the AI model V3, and the total value of the GPUs used is roughly RMB 385 million - 721 million yuan. DeepSeek explained that the development cost of V3 was $5.576 million; assuming AI training of 2.788 million hours at $2 per hour, this cost is less than one-tenth of US AI models. Professor Yutaka Matsuo of the University of Tokyo pointed out: developing an AI model requires dozens to hundreds of repeated experiments, and time was also spent before the ~2.8 million hours of learning, so this reasoning is logical — the time and GPUs spent on this should have been included in the cost. https://d9shhjt4p7ouc.cloudfront.net/china/ccompany/57993-2025-02-11-10-33-49.html
// - 5. how to "trust"? in survey form, were biases not excluded? no detailed quantification https://www.solidot.org/story?sid=80555
============
out-dated¶
https://x.com/wwwyesterday/status/1873785958891741412
AI may be fairly useful for solving front-end (or also some back-end) code, because the logic is relatively simple and the training data volume is large. But for non-front-end (e.g., C/C++ or perhaps other languages), because the training data scale is much smaller, it may not be as useful as the front-end (and in fact, even for the front-end, more specialized scenarios are not as useful as hyped online). Even if AI writes it, you still need to understand the code, fix bugs, and debug.
Some think estimating the overall code-line scale of a project is like a fairy tale. I can only feel regret. Of course, not every programmer needs this ability, but even as a small-team leader, if you have no sense of the order of magnitude (not necessarily precise) of the scale of the project (product) to be completed, that can only be called unqualified.
If you've been writing pretty much the same thing for 30 years (in other words, the thing is quite mature), then indeed you can estimate accurately; new things are impossible.
It's your own project, so you can estimate this way. The key is not the code, but the requirements.
AI requirements assessor?
Jan 06, 2025
============
https://polarisxu.studygolang.com/posts/basic/diagram-float-point/
https://www.zhihu.com/question/511856781/answer/2890338282
> Writing 96 million floating-point numbers to a file Their C++ code then took about 300 seconds Although I haven't used C++ for over ten years, I eventually optimized it to under 2 seconds
The project they had at the time involved scanning some simply-shaped parts, needing to fit the edge shape, length, thickness, etc., finally forming a file in a user-specified format. In subsequent collaboration, we found that the "algorithm" they described was a formula derived using knowledge of trigonometry and solid geometry. Anyone with a bit of programming sense understands that this is meaningless in real programming.
Although I had no prior knowledge of computer vision processing, I still learned OpenCV and solved some key problems, allowing the project to be delivered.
90% are people from mechanical or architectural backgrounds https://caadxyz.github.io/blog/python4rhino-zh
Either CV/CG algorithm engineers (people from school are most likely not) https://m.douban.com/book/subject/1841346/ Math / algorithms Not "computer algorithms or computer engineering"
============
Jan 6
The concept of information technology is very broad; those who need coders are not only the internet, but also traditional software, the ERP industry, automation, banking, etc. The knowledge depreciation rates in these industries differ. Banking still uses the antique COBOL, for many years now, and its obsolescence rate is extremely slow; automation/embedded C is almost irreplaceable; traditional software is more varied, mainly Java and other C-family languages, with a fairly fast knowledge depreciation rate, but still acceptable, because traditional software mostly targets specific customers, and mass-market products are a small share. The internet wants every product to face all 6 billion people globally, with a complex environment, many users, and extremely fast-changing requirements.
Although these types of products are all coders' work, the skill requirements are completely different. Banking needs data consistency, security, and can sacrifice speed; requirement changes are cautious, requiring a very high degree of business understanding. The industry entry barrier is not in technology but in business; banking has its own dedicated banking system testing.
Positions, as far as I know, generally only hire those with peer experience; it's not that you get hired just because your tech is awesome.
Traditional software needs a certain degree of changeability and deep understanding of business requirements. For example, for an aviation settlement system, you need to understand how settlement works; the frequency of requirement changes is not high. Traditional software has a role like "systems analyst," similar to a product manager but more technical, also considering the software's technical difficulties while balancing requirement changes.
The internet changes every day — today a campaign page launches, taken down overnight. Talking to everyone about systems analysis is nonsense; once it's done, whether the traffic can hold up without crashing, timeliness first, minor unimportant bugs are simply ignored; if the server goes down, as long as a restart fixes it, ship first and deal with it later (and a lot of "later" never gets followed up).
https://www.zhihu.com/question/41328565/answer/92389789 horizontal scaling https://x.com/ingramchen/status/1699739008723808629#m
static strong type , algebraic data type a devoted follower (of the school)
============
Jan 6
Using AI Generated Code Will Make You a Bad Programmer | Hacker News https://news.ycombinator.com/item?id=41913458
Front-end is "harder" because of "complex system" attributes, fast iteration
In the past there was no separation of front-end and back-end It was capital injection into the internet that gave rise to horizontal scaling (platform monopoly, traffic/attention economy)
Front-end technologies are not "progressing rapidly", they are thrashing. People are (rightfully) extremely frustrated with the limitations of the tools they have, so they (wrongfully) go off and invent new tools that get around those limitations, while inevitably re-introducing tons of difficulty in whatever the tools they're replacing were developed to avoid. https://news.ycombinator.com/item?id=37300579
Dealing with "complex systems" and "people" is thankless work
No money and PTSD on top: the worst job in the world https://www.36kr.com/p/1725124296705
If not for "faith and morality" https://www.zhihu.com/question/284312614/answer/465879747
Mathematics: an over-simplified "model" — having a sense of control prevents depression
============
Jan 6
People in compilers and reverse engineering (experts) will not be replaced
Recent work shows that LLMs demonstrate unequivocally superior results in being able to explain and resolve compiler error messages—for decades, one of the most frustrating parts of learning how to code. However, LLM-generated error message explanations have only been assessed by expert programmers in artificial conditions.
This work sought to understand how novice programmers resolve programming error messages (PEMs) in a more realistic scenario. We ran a within-subjects study with n = 106 participants in which students were tasked to fix six buggy C programs. For each program, participants were randomly assigned to fix the problem using either a stock compiler error message, an expert-handwritten error message, or an error message explanation generated by GPT-4.
Despite promising evidence on synthetic benchmarks, we found that GPT-4 generated error messages outperformed conventional compiler error messages in only 1 of the 6 tasks, measured by students’ time-to-fix each problem.
Handwritten explanations still outperform LLM and conventional error messages, both on objective and subjective measures. https://dl.acm.org/doi/fullHtml/10.1145/3689535.3689554 Not the Silver Bullet: LLM-enhanced Programming Error Messages are Ineffective in Practice
============
two major "tech" waves boosted with capitalism
- 1. web , front-end , cloud // attention-grasping economy & privacy surveillance : ad blocker , RSS & VPN , sino app
https://anvaka.github.io/map-of-github/#10.31/-19.0547/-34.1612
- 2. AI
============