Opus 4.7 uses an updated tokenizer that improves how the model processes text. The tradeoff is that the same input can map to more tokens—roughly 1.0–1.35× depending on the content type.

scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency.

After compressing, the model again extends its solutions to achieve stronger performance.

The standard autoresearch loop (brainstorm from code, run experiments, check metrics) works when the optimization surface is visible in the source. The Liquid results prove that. But for problems where the codebase doesn't contain enough information to generate good hypotheses, giving the agent access to papers and competing implementations changes what it tries.

The variance is also worth noting: baseline+FA TG has ±19 t/s of noise, while optimized+FA has ±0.59 t/s on x86. The fusions eliminate intermediate writes that pollute the cache, making the hot paths more predictable.

Without experience with compiler behavior, the agent couldn't have predicted which 'optimizations' the compiler would already handle.

A 606 MiB model at ~49 tokens/s consumes ~30 GB/s of memory bandwidth, close to the c6i.2xlarge's DRAM limit. No amount of SIMD tricks will help when the CPU is stalled waiting for model weights to arrive from DRAM.

Studying forks and other backends was more productive than searching arxiv. ik_llama.cpp and the CUDA backend directly informed two of the five final optimizations.

Coding agents working from code alone generate shallow hypotheses. Adding a research phase — arxiv papers, competing forks, other backends — produced 5 kernel fusions that made llama.cpp CPU inference 15% faster.

The variance is also worth noting: baseline+FA TG has ±19 t/s of noise, while optimized+FA has ±0.59 t/s on x86.

Total cost: ~$29 ($20 in CPU VMs, $9 in API calls) over ~3 hours with 4 VMs.

The agent fused them into one: for (int i = 0; i < nc; i++) { wp[i] = sp[i] * scale + mp_f32[i]; }

The same task on full Codex took ~5× longer.

KV Cache 内存占用降低 10.7 倍

GLM-5.1 pushes this frontier further, delivering 3.6× speedup and continuing to make progress well into the run. While its rate of improvement also slows over time, it sustains useful optimization for substantially longer than GLM-5.

GLM-5.1 did not plateau after 50 or 100 submissions, but continued to find meaningful improvements over 600+ iterations with 6,000+ tool calls, ultimately reaching 21.5k QPS—roughly 6× the best result achieved in a single 50-turn session.

In 23 months, the same capability that needed 1.8 trillion parameters now fits in 4 billion parameters. A 450x compression.

PagedAttention algorithm and vLLM system enhance the throughput of large language models by efficiently managing memory and reducing waste in the key-value cache.

Artificial Analysis has also positioned Gemini 3.1 Flash TTS within its 'most attractive quadrant' for its ideal blend of high-quality speech generation and low cost.

We introduce a pipelined double-buffered execution engine that overlaps parameter prefetching, computation, and gradient offloading across multiple CUDA streams, enabling continuous GPU execution.

identify if the app had been left in a broken state

the design of the retrieval and cache policy, especially how they decide what to keep, reuse, or drop across scenes, seems to be what actually drives the latency and throughput gains

Simultaneously, the parallelism in worker layer accelerates the speed of overall task execution, mitigating the significant latency

For higher-interactivity scenarios, execution time for MoE models is bound by expert weight load time. By splitting, or sharding, the experts across multiple GPUs across NVL72 nodes, this bottleneck is reduced, improving end-to-end performance.

Co-designed hardware, software, and models are key to delivering the highest AI factory throughput and lowest token cost. Measuring this goes far beyond peak chip specifications.

This means 2.7x more tokens from the same GB300 NVL72-based infrastructure and power footprint, reducing the cost to manufacture each token by more than 60%.

NVFP4 enables 4-bit precision while maintaining nearly identical accuracy to 8-bit precision, increasing performance per watt and lowering cost per token.

Optimizations that don’t need Rust Some of uv’s speed comes from Rust. But not as much as you’d think. Several key optimizations could be implemented in pip today: HTTP range requests for metadata. Wheel files are zip archives, and zip archives put their file listing at the end. uv tries PEP 658 metadata first, falls back to HTTP range requests for the zip central directory, then full wheel download, then building from source. Each step is slower and riskier. The design makes the fast path cover 99% of cases. This is HTTP protocol work, not Rust. Parallel downloads. pip downloads packages one at a time. uv downloads many at once. This is concurrency, not language magic. Global cache with hardlinks. pip copies packages into each virtual environment. uv keeps one copy globally and uses hardlinks (or copy-on-write on filesystems that support it). Installing the same package into ten venvs takes the same disk space as one. This is filesystem ops, not language-dependent. Python-free resolution. pip needs Python running to do anything, and invokes build backends as subprocesses to get metadata from legacy packages. uv parses TOML and wheel metadata natively, only spawning Python when it hits a setup.py-only package that has no other option. PubGrub resolver. uv uses the PubGrub algorithm, originally from Dart’s pub package manager. pip uses a backtracking resolver. PubGrub is faster at finding solutions and better at explaining failures. It’s an algorithm choice, not a language choice

No .egg support. Eggs were the pre-wheel binary format. pip still handles them; uv doesn’t even try. The format has been obsolete for over a decade. No pip.conf. uv ignores pip’s configuration files entirely. No parsing, no environment variable lookups, no inheritance from system-wide and per-user locations. No bytecode compilation by default. pip compiles .py files to .pyc during installation. uv skips this step, shaving time off every install. You can opt in if you want it. Virtual environments required. pip lets you install into system Python by default. uv inverts this, refusing to touch system Python without explicit flags. This removes a whole category of permission checks and safety code. Stricter spec enforcement. pip accepts malformed packages that technically violate packaging specs. uv rejects them. Less tolerance means less fallback logic. Ignoring requires-python upper bounds. When a package says it requires python<4.0, uv ignores the upper bound and only checks the lower. This reduces resolver backtracking dramatically since upper bounds are almost always wrong. Packages declare python<4.0 because they haven’t tested on Python 4, not because they’ll actually break. The constraint is defensive, not predictive. First-index wins by default. When multiple package indexes are configured, pip checks all of them. uv picks from the first index that has the package, stopping there. This prevents dependency confusion attacks and avoids extra network requests. Each of these is a code path pip has to execute and uv doesn’t.

PEP 658 went live on PyPI in May 2023. uv launched in February 2024. The timing isn’t coincidental. uv could be fast because the ecosystem finally had the infrastructure to support it. A tool like uv couldn’t have shipped in 2020. The standards weren’t there yet.

The similarity is because they are all saying roughly the same thing: Total (result) = Kinetic (cost) + Potential (benefit) Cost is either imaginary squared or negative (space-like), benefit is real (time-like), result is mass-like. Just like physics, the economic unfavourable models are the negative results. In economics, diversity of products is a strength as it allows better recovery from failure of any one, comically DEI of people fails miserably at this, because all people are not equal. Here are some other examples you will know if you do physics: E² + (ipc)² = (mc²)² (relativistic Einstein equation), mass being the result, energy time-like (potential), momentum the space-like (kinetic). ∇² - 1/c² ∂²/∂t² = (mc/ℏ)² (Klein-Gordon equation), mass is the result, ∂²/∂t² potential, ∇² is kinetic. Finally we have Dirac equation, which unlike the previous two as "sum of squares" is more like vector addition (first order differentials, not second). iℏγ⁰∂₀ψ + iℏγⁱ∂ᵢψ = mcψ First part is still the time-like potential, second part is the space-like kinetic, and the mass is still the result though all the same. This is because energy is all forms, when on a flat (free from outside influence) worksheet, acts just like a triangle between potential, kinetic and resultant energies. E.g. it is always of the form k² + p² = r², quite often kinetic is imaginary to potential (+,-,-,-) spacetime metric, quaternion mathematics. So the r² can be negative, or imaginary result if costs out way benefits, or work in is greater than work out. Useless but still mathematical solution. Just like physics, you always want the mass or result to be positive and real, or your going to lose energy to the surrounding field, with negative returns. Economic net loss do not last long, just like imaginary particles in physics.

Its performance is not very different from the system versions of grep, which shows that the recursive technique is not too costly and that it's not worth trying to tune the code.

Technique #1: Changing numeric representations

lossy compression: drop some of your data in a way that doesn’t impact your final results too much.

why not allow block forwarding without capturing: foo(&) foo(1, 2, &)

There is no reason why a small triangle couldn't be used in printing instead of the letter "e"

This is critical since many optimizations are accomplished by violating (hopefully safely) module boundaries; it is disastrous to incorporate optimizations into the main body of code. The separation allows the optimizations to be checked against the meanings.

If a compiler cannotdetermine whether or not two pointers may be aliased, it must assume that eithercase is possible, limiting the set of possible optimizations.

Algospeak refers to code words or turns of phrase users have adopted in an effort to create a brand-safe lexicon that will avoid getting their posts removed or down-ranked by content moderation systems. For instance, in many online videos, it’s common to say “unalive” rather than “dead,” “SA” instead of “sexual assault,” or “spicy eggplant” instead of “vibrator.”

It's typically taken for granted that better performance must require higher complexity. But I've often had the experience that making some component of a system faster allows the system as a whole to be simpler

The latest SQLite 3.8.7 alpha version is 50% faster than the 3.7.17 release from 16 months ago. [...] This is 50% faster at the low-level grunt work of moving bits on and off disk and search b-trees. We have achieved this by incorporating hundreds of micro-optimizations. Each micro-optimization might improve the performance by as little as 0.05%. If we get one that improves performance by 0.25%, that is considered a huge win. Each of these optimizations is unmeasurable on a real-world system (we have to use cachegrind to get repeatable run-times) but if you do enough of them, they add up.

don’t store all the values from onerow together, but store all the values from each column together instead. If each col‐umn is stored in a separate file, a query only needs to read and parse those columnsthat are used in that query, which can save a lot of work.

We were able to reduce calls to Infura and save on costs. And because the proxy uses an in-memory cache, we didn’t need to call the Ethereum blockchain every time.

our application continuously checks the current block and retrieves transactions from historical blocks;Our application was very heavy on reads and light on writes, so we thought using ‘cached reads’ would be a good approach;We developed a small application to act as a thin proxy between our application and Infura. The proxy application is simple and it only hits Infura/Ethereum on the initial call. All future calls for the same block or transaction are then returned from cache;Writes are automatically forwarded to Infura. This was a seamless optimization. We simply had to point our application to our proxy and everything worked without any changes to our application.

Grötschel, an expert in optimization, observes that a benchmark production planning model solved using linear programming would have taken 82 years to solve in 1988, using the computers and the linear programming algorithms of the day. Fifteen years later – in 2003 – this same model could be solved in roughly 1 minute, an improvement by a factor of roughly 43 million. Of this, a factor of roughly 1,000 was due to increased processor speed, whereas a factor of roughly 43,000 was due to improvements in algo-rithms

Optimization in this case is nothing crazy, just something I neglected while designing the framework.

If a UTF8-encoded Ruby string contains unicode characters, then indexing into that string becomes O(N). This can lead to very bad performance in string_end_with_semicolon?, as it would have to scan through the whole buffer for every single file. This commit fixes it to use UTF32 if there are any non-ascii characters in the files.

What is the point of avoiding the semicolon in concat_javascript_sources

   (intermediate value)(...) is not a function
       at something.source.js:1

   ({
     // other.js
   })()
   (function() {
     // something.js
   })();

And no need to walk backwards through all these strings which is surprisingly inefficient in Ruby.

Since the common problem with concatenating JavaScript files is the lack of semicolons, automatically adding one (that, like Sam said, will then be removed by the minifier if it's unnecessary) seems on the surface to be a perfectly fine speed optimization.

reducing it down to one call significantly speeds up the operation.

I feel like the walk in string_end_with_semicolon? is unnecessarily expensive when having an extra semicolon doesn't invalidate any JavaScript syntax.

Possibly use Array.prototype.filter.call instead of allocating a new array.

If you manage to make Svelte aware of what needs to be tracked, chances are that the resulting code will be more performant than if you roll your own with events or whatever. In part because it will use Svelte's runtime code that is already present in your app, in part because Svelte produces seriously optimized change tracking code, that would be hard to hand code all while keeping it human friendly. And in part because your change tracking targets will be more narrow.

In this study, we have quantitated yields of low copy and single copy number plasmid DNAs after growth of cells in four widely used broths (SB, SOC, TB, and 2xYT) and compared results to those obtained with LB, the most common E. coli cell growth medium

The template language's restrictions compared to JavaScript/JSX-built views are part of Svelte's performance story. It's able to optimize things ahead of time that are impossible with dynamic code because of the constraints. Here's a couple tweets from the author about that

GET is the primary mechanism of information retrieval and the focus of almost all performance optimizations.

It's fast. The Dart VM is highly optimized, and getting faster all the time (for the latest performance numbers, see perf.md). It's much faster than Ruby, and close to par with C++.

Note that when using sass (Dart Sass), synchronous compilation is twice as fast as asynchronous compilation by default, due to the overhead of asynchronous callbacks.

optcarrot, is a headless NES emulator that the Ruby core team are using as a CPU-intensive optimization target.

You might also want to check out the hyperxify browserify transform to statically compile hyperx into javascript expressions to save sending the hyperx parser down the wire.

JSX has the advantage of being fast, but the disadvantage that it needs to be preprocessed before working. By using template string virtual-html, we can have it work out of the box, and optimize it by writing a browserify transform. Best of both!

Static optimizations

I understand that I could use some third party memoization tool on top of the Svelte’s comparator, but my point here is — there is no magic pill, optimizations “out of the box” often turn out to have limitations.

This is a very dangerous practice as each optimization means making assumptions. If you are compressing an image you make an assumption that some payload can be cut out without seriously affecting the quality, if you are adding a cache to your backend you assume that the API will return same results. A correct assumption allows you to spare resources. A false assumption introduces a bug in your app. That’s why optimizations should be done consciously.

In the vast majority of cases there’s nothing wrong about wasted renders. They take so little resources that it is simply undetectable for a human eye. In fact, comparing each component’s props to its previous props shallowly (I’m not even talking about deeply) can be more resource extensive then simply re-rendering the entire subtree.

While you could use a map function for loops they aren't optimized.

"Premature optimization is the root of all evil"; start with RPC as default and later switch to REST or GraphQL when (and only if!) the need arises.

The more I think about this, the more I think that maybe React already has the right solution to this particular issue, and we're tying ourselves in knots trying to avoid unnecessary re-rendering. Basically, this JSX... <Foo {...a} b={1} {...c} d={2}/> ...translates to this JS: React.createElement(Foo, _extends({}, a, { b: 1 }, c, { d: 2 })); If we did the same thing (i.e. bail out of the optimisation allowed by knowing the attribute names ahead of time), our lives would get a lot simpler, and the performance characteristics would be pretty similar in all but somewhat contrived scenarios, I think. (It'll still be faster than React, anyway!)

By only adding the necessary handling when the function is declared async, the JavaScript engine can optimize your program for you.

The static analysis considerations make things like hero.enemies.map(...) a non-starter — the reason Svelte is able to beat most frameworks in benchmarks is that the compiler has a deep understanding of a component's structure, which becomes impossible when you allow constructs like that.

In some frameworks you may see recommendations to avoid inline event handlers for performance reasons, particularly inside loops. That advice doesn't apply to Svelte — the compiler will always do the right thing, whichever form you choose.

Performance optimizations are not free. They ALWAYS come with a cost but do NOT always come with a benefit to offset that cost.

Even so, the inline function is still created on every render, useCallback() just skips it.

Even having useCallback() returning the same function instance, it doesn’t bring any benefits because the optimization costs more than not having the optimization.

Optimization Models for Machine Learning: A Survey

Learning with Random Learning Rates

On the loss landscape of a class of deep neural networks with no bad local valleys

Revisiting Small Batch Training for Deep Neural Networks

Don't Use Large Mini-Batches, Use Local SGD

Accelerating Natural Gradient with Higher-Order Invariance

Backprop Evolution

Gradient Descent Finds Global Minima of Deep Neural Networks

A Convergence Theory for Deep Learning via Over-Parameterization

edge probes

For example, stochastic quasi-gradient algorithms [3] can be used forthe minimization of function (1).

While there are assets that have not been assigned to a cluster If only one asset remaining then Add a new cluster Only member is the remaining asset Else Find the asset with the Highest Average Correlation (HC) to all assets not yet been assigned to a Cluster Find the asset with the Lowest Average Correlation (LC) to all assets not yet assigned to a Cluster If Correlation between HC and LC > Threshold Add a new Cluster made of HC and LC Add to Cluster all other assets that have yet been assigned to a Cluster and have an Average Correlation to HC and LC > Threshold Else Add a Cluster made of HC Add to Cluster all other assets that have yet been assigned to a Cluster and have a Correlation to HC > Threshold Add a Cluster made of LC Add to Cluster all other assets that have yet been assigned to a Cluster and have Correlation to LC > Threshold End if End if End While

Effect of step size. The gradient tells us the direction in which the function has the steepest rate of increase, but it does not tell us how far along this direction we should step.

There are other ways of performing the optimization (e.g. LBFGS), but Gradient Descent is currently by far the most common and established way of optimizing Neural Network loss functions.