entrope · HackerLangs

10 saat öncediscuss

U2F was always an option, because I used it since they added the MFA requirement (and like others here never gave Github my phone number). I think TOTP was also available from the start. The warning to not use SMS for MFA might have been added later.

entrope

12 saat öncediscuss

Those are the ratios for memory bandwidth, but the GPUs have a much higher ratio for compute, and that affects prefill rate / TTFT, right?

entrope

evvelsi gündiscuss

Yes, all of those and more. Our measurement precision is much better than the year-to-year first and second derivatives of day length. https://datacenter.iers.org/singlePlot.php?plotname=Bulletin... has the most relevant plot to this; the vertical jumps reflect leap seconds. (IERS has other plots for other dimensions of rotation, but I like this one.)

entrope

evvelsi gündiscuss

IMO, "make your servers fungible" is a better way to express the intent: slightly shorter, no metaphors, although "fungible" is a less common word. Maybe that's just me. (Edited to add: "make your tests fungible" has I think the wrong connotation; I think the original wording on the blog is about test executions -- and "make your test executions fungible" does seem like a good goal, similar to ACID guarantees for database transactions.)

entrope

evvelsi gündiscuss

Can you elaborate on the distinction between "eliminate the hazard" as the first choice and "engineering out the hazard" as a fallback approach?

In my safety background (recently aerospace, ARP4754/4761), removing and avoiding the hazard are essentially equivalent, with reducing the likelihood and mitigating its effects acceptable if you can't remove or avoid the hazard, and procedure is also the least preferred mechanism.

entrope

5 gün öncediscuss

My experience is that AI agents write 20-33% more code than I would for a given feature set, mostly because they are worse at remembering what utility functions already exist and less likely to merge similar functions into more generic ones. They generate that code 2-10 times faster than I could. Defect density is harder to compare: I probably generate more "dumb" defects due to oversight or missing unit tests, but fewer defects that violate domain rules or architectural objectives.

entrope

5 gün öncediscuss

I think you're overlooking the lesson of Goodhart's law: you can use a metric, but if you make it a target, it stops being a good metric. Neither "tons of aircraft" nor "lines of code" should be the measuring stick -- and if they're not, then they can still be used as metrics.

To be fair, the hazard with AI agents is that they generate fluent output that is often facile, so it's easy to do a lot of things while having a lot of defects. That's a sign that quality control is not prioritized enough. A change in quality will also reduce the utility of SLOC as a metric, but the mechanism is different than what Charles Good hart pointed out.

entrope

11 gün öncediscuss

In this case, "Zero-JS" seems to mean that you write short bits of JavaScript that look like a domain-specific language, and you include a relatively small JavaScript file that translates certain server-sent events into DOM updates. The documentation for Datastar is less assertive; for example: > The value of the data-text attribute is a Datastar expression that is evaluated, meaning that we can use JavaScript in it. (from https://data-star.dev/guide/reactive_signals)

I personally would call this a minimal-JS approach to an SPA. The use of Nim as a server language and YottaDB as a storage engine seem incidental, except that the overall page is advertising for the database.

entrope

13 gün öncediscuss

A single benefit usually has an appropriate incentive structure, but a lot of people get multiple benefits -- even from different levels of government (local, state, federal) -- and adding up phase-outs in different systems can result in marginal phase-outs rates above 100%. It's hard to avoid that entirely given that we want to have a lot of transfers to the bottom of the income distribution while phasing those out by roughly the median. It would be easier to avoid phase-outs above (say) 80% of marginal income is we only had federal and state aid as predictable money transfers, but for various reasons we provide a lot of transfers in-kind or with limited authorized uses. Those limitations aren't necessarily wrong, but they do mean that transfers aren't fungible, so there's an incentive to provide transfers for other "good" uses, and that diversity is what makes it hard to bound the marginal phase-outs for everyone.

entrope

15 gün öncediscuss

Perhaps to support demand for the products with recent price hikes, and/or the upcoming Mac Studio with M5 Ultra, rather than have customers sit on the sidelines thinking they'll wait this generation out.

I am still skeptical that Apple intentionally leaked this because they normally are so tight-lipped, but there are reasons in favor of leaking this.

entrope

18 gün öncediscuss

The mechanism is that generating tokens (the "decode" phase) in an LLM is limited by memory bandwidth for the weights, so computing multiple streams amortizes the bandwidth over streams as long as you can keep the contexts in RAM. This is most true for dense models and the always-on expert in MoE models, or when you have significantly more streams than experts for MoE models.

In contrast, prompt prefill is more easily compute-bound, so there are interesting trade-offs for latency of decode vs prefill when the LLM utilization is high.

entrope

18 gün öncediscuss

> The DGX Spark has basically the same memory bandwidth as a M5 Pro, and far more than a M5.

I see ~274 GB/sec for the DGX Spark[1], versus 307 GB/sec for M5 Pro and 460 or 614 GB/sec for M5 Max[2]. One might call 90% "basically the same", but there are nominally two tiers above "Pro".

Yes, a MacBook Pro with 128 GB and M5 Max costs $5100 (14") or $5400 (16") versus currently $4700 for the DGX Spark, but the MBP includes keyboard, mouse, battery and portability. I believe its prefill is slower and you get 2 TB vs 4 TB SSD, but overall one gives up a lot to save 10% of the cost.

[1]- https://docs.nvidia.com/dgx/dgx-spark/hardware.html [2]- https://support.apple.com/en-us/126319

entrope

19 gün öncediscuss

For coding, Qwen3.6-27B with MTP should fit in 32GB with almost full context length for Unsloth's 5-bit quantization. That's my preferred choice for a local coding agent on similar hardware: the quality delta compared to a MoE model is IMO worth the extra wait. (And I haven't found a model with 70B-120B parameters that works better for coding.) For general chat, maybe gpt-oss-120b? It should have more general knowledge than a 30B-class model; I've used it to suggest itineraries for trips and to review the completeness of small requests for proposals.

I don't have recommendations for images because I haven't played with those.

entrope

24 gün öncediscuss

HuggingFace says this model has 753B parameters, which will need a lot more RAM than a maxed-out MacBook Pro. With 40B active parameters, running from SSD would need patience.

entrope

28 gün öncediscuss

The answer to the question posed in the site's domain name is "no", unfortunately.

It looks like it just grabbed the intro to each project's self-description, but blurbs like "Zero-cost ultra-high-performance declarative DOM library using FRP signals" would be worth very little even with screenshots.

entrope

28 gün öncediscuss

Ignoring multiple evaluation, one can also #define stdc_has_single_bit(X) !((X) & ((X)-1)). If X isn't a power of two, the -1 will leave the MSB in place.

entrope

28 gün öncediscuss

Yes, but presumably the authors are suggesting broader application than just caching a system prompt.

The paper's approach should work well if (a) you can calculate KV(A || B) as a function of KV(A) and KV(B) independently, (b) you can identify which documents A1, A2, A3, ... are used commonly enough to be worth caching, and (c) it is cheaper to buy and sell KV(A) on a market than to compute KV(A) when it is needed. Given the size of KV(A) I am not sure that (c) will become true even if people solve the open research problem represented by (a) and accept the state-of-the-art trade-offs known for (b).

entrope

geçen aydiscuss

Perhaps I overlooked something, but which part of this involved "fooling Go"? I would usually not call it "fooling" something to trigger a not-strictly-required rejection of a dubious trait, especially when best practice says to avoid that trait.

entrope

geçen aydiscuss

And the Russian system is named (the Russian words for) "Global Navigation Satellite System", but usually only called GLONASS because adding L, O and A is less confusing than having one name for super- and sub-sets in a single category.

The fourth global GNSS constellation is Europe's Galileo. NavIC and QZSS are regional GNSS constellations.

entrope

geçen aydiscuss

Two things amaze me about GPS. First, that there are still four Block IIR and seven Block IIR-M satellites operational; these had 7.5 year design lives and were launched by 2004 and 2009 respectively. Second, that L1C, L5 and L2C are all still pre-operational thanks to the OCX debacle. L1C and L2C really modernize the signal structure to improve accuracy.

comments