HNHacker News
TopNewBestAskShowJobs

Majromax

3,557 karma · joined February 1, 2019

submissionscomments
Majromax··on Measuring Claude 4.7's tokenizer costs
> Similarly, it's unlikely you can measure a significant performance difference between models like GPT 5.4-xhigh and GPT 5.2 unless you have a task where one of them almost always fails or one almost always succeeds

That feels like a concession to the limited benchmarking framework. 5.4-xhigh is supposed to be (and is widely believe to be) a better model than 5.2, so if that's invisible in the benchmarking scores then the protocol has problems. The test probably should include cases that should be 'easy passes' or 'near always failures', and then paired testing could offer greater precision on improvements or degradations.

Conversely, if model providers also don't do this then they could be accidentally 'benchmaxxing' if they use protocols like this to set dynamic quantization levels for inference. All you really need for a credible observation of problems from 'less intensive use' is a problem domain that isn't well-covered by the measured and monitored benchmark.

Majromax··on Guy builds AI driven hardware hacker arm from duct tape, old cam and CNC machine
Decoupling capacitors don't remove the ground reference, they just allow high-frequency signals a faster path to ground.

Typically, you need dedicated circuitry (and usually inductive coupling) to provide full isolation, but if the circuit is using this layout then you can still choose to ground the normally-isolated side for probing.

Majromax··on Measuring Claude 4.7's tokenizer costs
Caveman doesn't and cannot change the tokenizer, so the relative token count differences by input category will remain unchanged.
Majromax··on Claude Opus 4.7 costs 20–30% more per session
While that's a nice effort, the inter-run variability is too high to diagnose anything short of catastrophic model degradation. The typical 95% confidence interval runs from 35% to 65% pass rates, a full factor of two performance difference.

Moreover, on the companion codex graphs (https://marginlab.ai/trackers/codex-historical-performance/), you can see a few different GPT model releases marked yet none correspond to a visual break in the series. Either GPT 5.4-xhigh is no more powerful than GPT 5.2, or the benchmarking apparatus is not sensitive enough to detect such changes.

Majromax··on Claude Opus 4.7
Don't confuse the many voices of a crowd with a single person's fickle view. If you can track an individual person or organization who changes their mind 'every other week' then more power to you, but unless you're performing that longitudinal study you are simply seeing differential levels of enthusiasm.
Majromax··on Introspective Diffusion Language Models
Are Macs/etc compute bound with their 'it fits in unified memory' language models? Certainly by the time you're streaming weights from SSD you must be back in a bandwidth-bound regime.
Majromax··on Backblaze has stopped backing up OneDrive and Dropbox folders and maybe others
You meant “reasonable,” but you did not apply reason. Situations such as this can be handled with a quota set at something like 150% of median use, but then extended upon a justified request. It can work in a lab where there’s a human touch, but it fails at million-user scale where even that level of human support is too expensive.
Majromax··on Nothing Ever Happens: Polymarket bot that always buys No on non-sports markets
That's just the binary nature of the bet. You address that in a real trading strategy with (fractional) Kelly position sizes. Anyone doing this for actual money would also be well served by implementing continuous monitoring and active risk management over top, in order to limit maximum drawdowns if the trend evaporates.
Majromax··on Nothing Ever Happens: Polymarket bot that always buys No on non-sports markets
> picking up pennies in front of a steamroller.

This or any other statistical play is only 'in front of a steamroller' if you do it with leverage, especially if notionally uncorollated bets suddenly move together. Bets on Polymarket have limited downside by design, and bets in different categories are obviously unrelated to each other.

Without having looked into this in detail, however, I suspect the problem would be limited capacity; markets that are both deep and so trivially irrational are probably fairly infrequent. You might pick up pennies but only pennies.

Majromax··on Nothing Ever Happens: Polymarket bot that always buys No on non-sports markets
> Bookmakers price purely based on facts and statistics. Their pricing isn't affected by excitement nor by how many people are betting a certain way.

A bookmaker is a market maker, and they ideally want to end up with no net interest in a position. They then take guaranteed profit in the bid-ask spread, which in sportsbooks is the 'vig'. Bookmakers who adjust their odds in real-time don't have to be particularly clever about the fundamentals, just responsive to the competing demands on either side.

A bookmaker who intentionally takes a position on a game is the equivalent of a proprietary trader or hedge fund. It's potentially more profitable, but it's also adversarial against 'sharp' traders.

Bookmakers who set odds at the beginning and don't move with the action must set larger bid-ask spreads to compensate.

Majromax··on Nothing Ever Happens: Polymarket bot that always buys No on non-sports markets
> Edit: conversely, if the average no costs _more_ than 73 cents, but the 73% of all polymarkets resolve to No, that would imply that an everything-always-happens strategy is profitable (neglecting slippage)

Or just the bid-ask spread; price no at 73.25 and yes at 27.5 and you have a profitable but theoretical mid-market price.

Majromax··on Nothing Ever Happens: Polymarket bot that always buys No on non-sports markets
'Enough' [capital] is doing a lot of work in that sentence. In the limit of a one-sided irrational market, the 'rational actor' would need to take the other side of every open transaction.
Majromax··on Nothing Ever Happens: Polymarket bot that always buys No on non-sports markets
Wet streets cause rain? You don’t show that the Polymarket signal lead Nasdaq.
Majromax··on Industrial design files for Keychron keyboards and mice
> Not a lawyer, but as I understand it the license is a matter of copyright, and the copyright only applies to the design files. So as long as you're making that keyboard for yourself then you should be good to do anything you want with the keyboard, because it is no longer using the license at that point.

What if I take the design, print it, include the thing in a staged photo, and sell prints of the photo?

What if I skip the printing and use the design files as a basis for a rendered photo or animation?

What if I print the design, then use a 3D scanner to recreate a file from the physical artifact?

Majromax··on The Training Example Lie Bracket
Well, the "vector field defined by the update attributable to this training sample" is a well-defined thing (even if it's not just the gradient of loss with respect to parameters), so that part translates.

However, what's harder to interpret is how this field transports with respect to θ, since the momentum vector and θ are themselves inextricably linked. If you somehow arrived at a different θ, then you'd have a different momentum. (On the gripping hand, the bracket is a construct of infinitesimals, maybe that doesn't matter.)

Majromax··on I still prefer MCP over skills
Anyone who's ever `DROP TABLE`d on a production rather than test database has encountered the same problem in meatspace.

In this context, the MCP interface acts as a privilege-limiting proxy between the actor (LLM/agent) and the tool, and it's little different from the standard best practice of always using accounts (and API keys) with the minimum set of necessary privileges.

It might be easier in practice to set up an MCP server to do this privilege-limiting than to refactor an API or CLI-tool, but that's more an indictment of the latter than an endorsement of the former.

Majromax··on The Training Example Lie Bracket
Wait a second, they define the induced vector field (and consequently Lie bracket) in terms of batch-size 1 SGD:

> In particular, if x is a training example and L(x) is the per-example loss for the training example x, then this vector field is: v^(x)(θ) = -∇_θ L(x). In other words, for a specific training example, the arrows of the resulting vector field point in the direction that the parameters should be updated.

but for the MXResNet example:

> The optimizer is Adam, with the following parameters: lr = 5e-3, betas = (0.8, 0.999)

This changes the direction of the updates, such that I'm not completely sure the intuitive equivalence holds.

If it were just SGD with momentum, then the measured update directions would be a combination of the momentum vector and v1/v2, so {M + v1, M + v2} = {v1, M} + {M, v2} + {v1, v2}. The Lie bracket is no longer "just" a function of the model parameters and the training examples; it's now inherently path dependent.

For Adam, the parameter-wise normalization by the second norm will also slightly change the directions of the updates in a nonlinear way (thanks to the β2 term).

The interpretation is also strained with fancier optimizers like Muon; this uses both momentum and (approximate) SVD normalization, so I'm really not sure what to expect.

Majromax··on Six (and a half) intuitions for KL divergence
The math doesn't need a 'true' or 'false' distribution; that just falls out of the use of a model ('false') to approximate reality ('true'). When the bard says "there are more things in heaven and earth, Horatio, than are dreamt of in your philosophy," he's also saying that the KL Divergence between Horatio's beliefs and reality is infinite.

We can also apply the concept between two subjective distributions. If I'm indifferent to sports teams (very broad distribution) and you're a rabid fan of A (sharp, narrow distribution), then it might take you a long time to express a point in a way I'll understand – but conversely I might be able to express "team B is good actually" in a way that just does not compute for you.

Majromax··on Škoda DuoBell: A bicycle bell that penetrates noise-cancelling headphones
There's also another difference: beeps can reflect coherently off of surfaces, causing directionality confusion in a dense environment. White noise is much less likely to have odd interference patterns, maximizing our ability to localize the sound.
Majromax··on We found an undocumented bug in the Apollo 11 guidance computer code
Since the advantage of standards is that there are so many to choose from, one lesser-used but still regionally acceptable approach (e.g. https://www.alberta.ca/web-writing-style-guide-punctuation#j...) is to use en-dashes offset with spaces.
Majromax··on Germany Power Prices Turn Deeply Negative on Renewables Surge
If you're buying and selling the power, then you don't care much about the absolute price, but rather the delta between the purchase and sale prices.

Below -100€/MWh, you don't need to sell the power to profit; you'd make at least some money just using it to heat up a big resistor.

Majromax··on Issue: Claude Code is unusable for complex engineering tasks with Feb updates
I'm curious about your subscription/API comparison with respect to thinking. Do you have a benchmark for this, where the same set of prompts under a Claude Code subscription result in significantly different levels of effective thinking effort compared to a Claude Code+API call?

Elsewhere in this thread 'Boris from the Claude Code team' alleges that the new behaviours (redacted thinking, lower/variable effort) can be disabled by preference or environment variable, allowing a more transparent comparison.

Majromax··on Issue: Claude Code is unusable for complex engineering tasks with Feb updates
If I'm reading that page correctly, then the benchmark results don't cover the interesting "mid February" inflection point noted in the article/report. The numbers appear to begin after the quality drop began. Moreover, the daily confidence interval seems to be stupidly wide, with a confidence interval between 42% and 69%?

The "Other metrics" graphs extend for a longer period, and those do seem to correlate with the report. Notably, the 'input tokens' (and consequently API cost) roughly halve (from 120M to 60M) between the beginning of February and mid-March, while the number of output tokens remains similar. That's consistent with the report's observation that new!Opus is more eager to edit code and skips reading/research steps.

Majromax··on Issue: Claude Code is unusable for complex engineering tasks with Feb updates
If you're using API pricing, then you can bring your own harness with full visibility/oversight of the prompting.
Majromax··on Costco sued for seeking refunds on tariffs customers paid
> this is a wild fantasy when dealing with any commercial entity with a fiduciary responsibility to shareholders.

"Fiduciary duty" is less strict than you'd expect. Courts generally recognize a "business judgment rule," where executives are offered broad discretion in strategy subject to some basic reasonability tests.

This would allow Costco to say "in order to cultivate goodwill and maintain our reputation, after we receive refunds we will distribute them to our customers based on purchased goods with refunded tariffs." It would also allow the directors to book the refund as profits, or use it for later incentives or marketing, or a variety of other actions.

The 'fiduciary duty' aspect here is mostly a myth. Directors do indeed have a fiduciary duty, but that duty is towards the corporation as a whole – including its long-term interests – rather than strictly towards short-term profit maximization. The fiduciary duty doctrine exists more to prevent graft and self-dealing, where managers and directors 'loot' the company by smuggling out profits in ways that benefit themselves personally rather than the company as a whole.

Majromax··on Data centers are transitioning from AC to DC
With modern technologies, that's power over ethernet or USB-C. Other comments in this thread point out that the telephone service also routinely used 48V for the ring signal.

However, higher DC voltage is riskier, and it's not at all standard for electrical and building code reasons. In particular, breaking DC circuits is more difficult because there's no zero-crossing point to naturally extinguish an arc, and 170V (US/120VAC) or 340V (Europe/240VAC) is enough to start a substantial arc under the right circumstances.

Unfortunately for your lighting, it's also both simple and efficient to stack enough LEDs together such that their forward voltage drop is approximately the rectified peak (i.e. targeting that 170/340V peak). That means that the bulb needs only one serial string of LEDs without parallel balancing, making the rest of the circuitry (including voltage regulation, which would still be necessary in DC world) simpler.

Majromax··on Arm AGI CPU
> My definition of AGI hasn't changed - it's something that can perform, or learn to perform, any intellectual task that a human can.

Wait, could you make your qualifiers specific here? Is your definition of AGI that it be able to perform/learn any intellectual task that is achievable by every human, or by any human?

Those are almost incomparably different standards. For the first, a nascent AGI would only need to perform a bit better than a "profound intellectual disability" level. For the second, AGI would need to be a real "Renaissance AGI," capable of advancing the frontiers of thought in every discipline, but at the same time every human would likely fail that bar.

Majromax··on Cursor Composer 2 is just Kimi K2.5 with RL
> this may also be in fact enough evidence for Anthropic to ban Cursor from using their models, just like they are doing to OpenCode.

The Anthropic ban on OpenCode isn't an Anthropic ban on OpenCode, it's a ban on using a Calude Code subscription with OpenCode. That's justified (or not) under various ToS arguments, but one can still use OpenCode with the more expensive API access.

Anthropic's complaint about distillation attacks is a distinct prong, one not levied against OpenCode. Additionally, the distillation activities described in your link don't describe Cursor's routine use of Anthropic's models. There, the model outputs are a primary product (e.g. the autocompleted code), and any learning signals provided are incidental.

Majromax··on Cursor Composer 2 is just Kimi K2.5 with RL
> The worthwhile question AIUI is whether AI weights are even protected by human copyright.

I'm also deeply curious about this legal question.

As I see it, model weights are the result of a mechanistic and lossy translation between training data and the final output weights. There is some human creativity involved, but that creativity is found exclusively in the model's code and training data, which are independently covered by copyright. Training is like a very expensive compilation process, and we have long-established that compiled artifacts are not distinct acts of creation.

In the case of a proprietary model like Kimi, copyright might survive based on 'special sauce' training like reinforcement learning – although that competes against the argument that pretraining on copyrighted data is 'fair use' transformation. However, I can't see a good argument that a model trained on a fully public domain dataset (with a genuinely open-source architecture) could support a copyright claim.

Majromax··on No, it doesn't cost Anthropic $5k per Claude Code user
> 1) Anthropic doesn't want their subsidised plans being used outside of CC, which would imply that the money their making off it isn't enough, a

Claude Code use-cases also differ somewhat from general API use, where the former is engineered for high cache utilization. We know from overall API costs (both Anthropic and OpenRouter) that cached inputs cost an order of magnitude less than uncached inputs, but OpenCode/pi/OpenClaw don't necessarily have the same kind of aggressive cache-use optimizations.

Vertically integrated stacks might also be able to have a first layer of globally shared KV cache for the system prompts, if the preamble is not user specific and changes rarely.

> 2) last time I checked, API spending is capped at $5000 a month

Per https://platform.claude.com/docs/en/api/rate-limits, that seems to only be true for general credit-funded accounts. If you contact Anthropic's sales team and set up monthly invoicing, there's evidently no fixed spending limit.

← PreviousPage 3 of 27Next →