claiir

The writing of the paper was so heavily ai, I was unsure if it was satire given the subject matter.

currently (as of Q3 2025) the fastest-growing of the top four languages in the world… +90% users in the past 3.5 years.

Because of AI, right?

Oh this makes sense.

And Gemini 3 can’t..? Isn’t this just a thinking vs nonthinking model thing?

This part is interesting:

[verify to] Speak in a stage channel.

My understanding is non-stage voice channels are E2E encrypted, and Discord retains no recordings, whereas stage channels are not. Is this a liability thing—Discord not wanting to have voice recordings of non-adults?

Javascript doesn’t generally execute on your GPU.

Asked by CNN's Kaitlan Collins why previous editions of the National Climate Assessment were no longer available online, former fracking company CEO Wright responded [..]

lol

Maybe just have commands auto-execute if you click on links in the existing text? That would allow someone to experience the entire interface on a touch device! :>

E.g. there is **__contact__** in the page, bold and underlined, but you cannot click on it to do anything.

Sounds like the LLM you used when writing this slop comment struggled with the problem too. :>

Same experience with my personal benchmarks. Generally unimpressed with Qwen3.

o1-preview had this same issue too! You’d give it a long conversation to summarize, and if the conversation ended with a question, o1-preview would answer that, completely ignoring your instructions.

Generally unimpressed with Qwen3 from my own personal set of problems.

Likely not the case, given (1) the body was peri-mortem decapitated (by a human) and (2) apparent structural damage was limited to a single bite mark (on the ilium), with no signs of "taphonomic" damage (indicating limited soft tissue trauma)? [1]

(1) > 6DT19 had been decapitated with a single cut between the second and third cervical vertebrae , delivered from behind.

(2) > Additional [to the decapitation] peri-mortem trauma was present in the form of a series of small depressions on both sides of the pelvis [..]

Taphonomic damage alone is also unlikely due to the appearance and margins of the lesions, which are the same colour as the surrounding bone (this differs if the break is post-mortem; [56]), and the adherence of bony fragments at the injury site (which occurs when soft tissue is present) .

[1] https://journals.plos.org/plosone/article?id=10.1371/journal...

Actual paper: https://journals.plos.org/plosone/article?id=10.1371/journal...

GoDaddy is actively experimenting to integrate image generation so customers can easily create logos that are editable [..]

I remember meeting someone on Discord 1-2 years ago (?) working on a GoDaddy effort to have customer-generated icons using bespoke foundation image gen models? Suppose that kind of bespoke model at that scale is ripe for replacement by gpt-image-1, given the instruction-following ability / steerability?

Okay, but why did `LeaveCriticalSection` change? Compiler changes, new features, refactoring, etc? That’s the most interesting part (and absent)!

It’s not just you. The speedup is an artefact of the CFG (Classifier-Free Guidance) the model uses. The other problem is the speedup isn’t constant—it actually accelerates as the generation progresses. The Parakeet paper [1] (which OP lifted their model architecture almost directly from [2]) gives a fairly robust treatment to the matter:

When we apply CFG to Parakeet sampling, quality is significantly improved. However, on inspecting generations, there tends to be a dramatic speed-up over the duration of the sample (i.e. the rate of speaking increases significantly over time). Our intuition for this problem is as follows: Say that is our model is (at some level) predicting phonemes and the ground truth distribution for the next phoneme occuring is 25% at a given timestep. Our conditional model may predict 20%, but because our uncondtional model cannot see the text transcription, its prediction for the correct next phoneme will be much lower, say 5%. With a reasonable level of CFG, because [the logit delta] will be large for the correct next phoneme, we’ll obtain a much higher final probability, say 50%, which biases our generation towards faster speech. [emphasis mine]

Parakeet details a solution to this, though this was not adopted (yet?) by Dia:

To address this, we introduce CFG-filter, a modification to CFG that mitigates the speed drift. The idea is to first apply the CFG calculation to obtain a new set of logits as before, but rather than use these logits to sample, we use these logits to obtain a top-k mask to apply to our original conditional logits. Intuitively, this serves to constrict the space of possible “phonemes” to text-aligned phonemes without heavily biasing the relative probabilities of these phonemes (or for example, start next word vs pause more). [emphasis mine]

The paper contains audio samples with ablations you can listen to.

[1] https://jordandarefsky.com/blog/2024/parakeet/#classifier-fr...

[2] https://news.ycombinator.com/item?id=43758686

Yea they mention a “perplexity drop” relative to naive quantization, but that’s meaningless to me. > We reduce the perplexity drop by 54% (using llama.cpp perplexity evaluation) when quantizing down to Q4_0.

Wish they showed benchmarks / added quantized versions to the arena! :>

In this case it's a little bit worse; the "nate" app had a literally "0% automation rate," despite representations to investors of an "AI" automation rate of "93-97%" powered by "LSTMs, NLP, and RL." No ML model ever existed! [1]

See:

As SANIGER knew, at the time nate was claiming to use AI to automate online purchases, the app’s actual automation rate was effectively 0%. SANIGER concealed that reality from investors and most nate employees: he told employees to keep nate’s automation rate secret; he restricted access to nate’s “automation rate dashboard,” which displayed automation metrics; and he provided false explanations for his secrecy, such as the automation data was a “trade secret.”

SANIGER claimed that nate's "deep learning models" were "custom built" and use a "mix of long short-term memory, natural language processing, and reinforcement learning."

When, on the eve of making an investment, an employee of Investment Firm-1 asked SANIGER about nate's automation rate, that is, the percentage of transactions successfully completed with nate's AI technology, SANIGER claimed that internal testing showed that "success ranges from 93% to 97%."

(from [1])

[1] https://www.justice.gov/usao-sdny/media/1396131/dl?inline

Actual indictment: https://www.justice.gov/usao-sdny/media/1396131/dl?inline

* Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith.

Be charitable. I provided encyclopedic reference out of a mutual discovery interest. Magazine articles[/opeds] may—in your words—present “opinions of some,” whereas reference-grade material provides a broad and citable foundation—ostensibly what you are looking for (“your axiomatic foundation is just another opinion of some”).

You haven't [..] backed up your opinion

In the spirit of “respond to the strongest plausible interpretation” and good faith, I think you may have missed the argument I was making above. :>

I am positing a reformulation/distillation of “positive trade balance preference” as “preference for foreign investment,” drawing on Palgrave, although perhaps controversially. The former is controversial, especially if seen as a-priori; the latter mundane.

:>

The fundamental precept here is that foreign investment of goods is more preferable to domestic investment of goods [1], so running a negative trade balance is dispreferable.

Trump is a textbook Mercantilist (positive trade balance inherently good [2]). Rather than magazine articles, see the Palgrave (the definitive encyclopedic reference for Econ) entry [1] and the Wikipedia page [2] for the arguments and history.

I always find reading the highest quality material leads to the highest quality thinking—there’s certainly a reason why modern LLM training mixtures might weight wikipedia tokens 5x and webtext 0.5x. :>

[1] https://link.springer.com/referenceworkentry/10.1057/978-1-3...

[2] https://en.wikipedia.org/wiki/Mercantilism

Long-term trade deficits are intrinsically bad. The tariffs directly address this by being exactly weighed by said deficit. See the OTR report [1]:

To conceptualize reciprocal tariffs, the tariff rates that would drive bilateral trade deficits to zero were computed. While models of international trade generally assume that trade will balance itself over time, the United States has run persistent current account deficits for five decades, indicating that the core premise of most trade models is incorrect.

The failure of trade deficits to balance has many causes, with tariff and non-tariff economic fundamentals as major contributors. Regulatory barriers to American products, environmental reviews, differences in consumption tax rates, compliance hurdles and costs, currency manipulation and undervaluation all serve to deter American goods and keep trade balances distorted. As a result, U.S. consumer demand has been siphoned out of the U.S. economy into the global economy, leading to the closure of more than 90,000 American factories since 1997, and a decline in our manufacturing workforce of more than 6.6 million jobs, more than a third from its peak.

While individually computing the trade deficit effects of tens of thousands of tariff, regulatory, tax and other policies in each country is complex, if not impossible, their combined effects can be proxied by computing the tariff level consistent with driving bilateral trade deficits to zero. If trade deficits are persistent because of tariff and non-tariff policies and fundamentals, then the tariff rate consistent with offsetting these policies and fundamentals is reciprocal and fair.

[1] https://ustr.gov/issue-areas/reciprocal-tariff-calculations

Tariffs are not a lever, they distort risk tolerance.

The "distortion" / divergence from market trends is exactly the point. This is no different from the FED setting the risk-free interest rate, or a carbon tax being levied. The government is holding actors accountable for an externality (trade deficit).

Tariffs are crude, they are like a huge hammer hitting everyone, including allies

The intrinsically bad part is the trade imbalance, not adversarial nation status; these tariffs are structured such that if adhered to, ceteris paribus, we will have zero trade imbalance. See the OTR report. These tariffs are anything but crude, rather calculated and surgical. Countries with little deficit will be changed little tariff; those with large deficit a proportional, calculated tariff.

Tariffs cause loss of confidence and loss of credit

This has yet to be shown in the long term.

Blanket tariffs target allies; they're isolating us and we're creating blocks polarized to us and weaning themselves off our economic output.

Again, these are not "blanket tariffs." They are strategic, surgical and calculated, based directly on the particular trade deficit--especially indicated for allies who have gone unchecked for far too long. It's unclear if adversarial BLOCs have much to do with tariffs, and the US's economic exports (especially key exports like oil and LNG) have only increased amidst recently increasing strategic import controls.

Tariffs only reduce dependency if there is internal capacity, of which there is none.

You're putting the cart before the horse. We need demand before we see an increase in production.

We'll best case shift our dependency via third-party countries We're concentrating our risk on fewer countries where tariffs are lower

Reciprocal tariffs are calculated and recalculated based on current trade deficit--this won't happen. Intrinsically, the "bucketing" of trade along geopolitical borders (we sum all exports and imports along national lines), actually incentives nation-heterogeneity, not nation-homogeneity, in imports, which is directly in line with national security interest.

We're reducing our efficiency and increasing prices, job losses are coming, productivity will stagnate. We're reducing our agility and our headroom

Prices will increase, naturally. That's the entire point. See the OTR report. Jobs likely won't decrease, though, given the cost of labor will be relatively cheaper than the cost of goods, incentivizing hiring. Additionally, the FED is likely to lower interest rates, which will put further demand on the labor market. :> We certainly have a lot more negotiating power and agility and headroom now that we can strategically peel away tariffs to absorb supply shocks, once the economy adjusts to the new status quo.

Historically, autarky has lead to stagnation.

When has the US ever been in autarky?

Ultimately, like the FED setting the risk-free interest rates, the POTUS is tempering the market's risk-taking behavior by internalizing the externality of trade deficit for economic agents.

In the long-term. Hence "innoculation"--a prick today to prevent a serious illness tomorrow.

In the grand scheme of things, the market movement we saw in response to these tariffs is dwarfed by the sustained market downturn throughout much of 2022. Fueled by persistent inflation concerns and the FED's aggressive interest rate hikes, the S&P 500 index shed approximately 25% from its peak in January 2022 to its low in October 2022, a far more significant and prolonged decline than a single day's reaction to tariff news.

We recovered from those. We will recover from the short-term effects of these tariffs as our economy begins to price in the risk of external dependency.

Correct; it doesn't. However,

(1) Tariffs will not affect all commodities in the same class equally, creating a natural price advantage, even at the level of consumers. While not absolute autarky, demand will continue to shift to less externally-dependent goods, which has and will continue to reduce our liability.

(2) Ideally congress would codify the tariffs to prevent that. That said, it seems the first Trump term tariffs did induce production onshoring, despite that uncertainty.

What would it take to change your mind on these tariffs?

Original clip is here [1]. This likely isn't real, though if ChatGPT were used, I suppose it worked out this time :>

[1] https://www.tiktok.com/@destinygnome/video/74891340165247337...

Thanks for the question. Three reasons:

(1) Much like a carbon tax, when the economy becomes accountable for an externality (risk in this case, rather than carbon) it is priced in at every level, ending with consumers, whose demand is price-sensitive. Not all consumer goods of the same class will be impacted equally, creating a natural price advantage.

(2) Once the economy adjusts to these new risk-adjusted prices, POTUS can always offer clemency for specific goods and industries to respond to global supply shocks, providing a large buffer for the now-adjusted American consumer.

(3) Narrowly: it is exceedingly unlikely we won't see production onshoring in any capacity like you suggest. Economic analysis of the first-term Trump tariffs, for example, did indeed find production onshoring.

I am in the minority here. These tariffs are fairly reasonable. Here's why:

(1) The state has a compelling interest in regulating market risk tolerance. When the FED sets the risk-free interest rate, it controls the price of risk, and thereby the risk of the entire market. It is often advisable to temper the risk-eager market for long-term stability.

(2) Likewise, when the POTUS enacts tariffs, it is similarly setting the market's risk tolerance--namely against the risks concomitant of economic dependency on foreign nations.

(3) This will inoculate the economy against external supply chain shocks (at the extreme: war) by reducing our liability. It's not just consumers who are price-sensitive; given goods within a class will not be affected equally, natural price advantages will emerge, shifting demand and consumption as all levels of economy begin to price-in these tariffs.

The "economic consensus" against these tariffs just isn't real? It's exceedingly likely that the US likely will be better off because of them :)

(2) is also significantly harder with these new models as they don’t support word timestamps like WHISPR.

see > Other parameters, such as timestamp_granularities, require verbose_json output and are therefore only available when using whisper-1.

Hi Jeff. This is awesome. Any plans to add word timestamps to the new speech-to-text models, though?

Other parameters, such as timestamp_granularities, require verbose_json output and are therefore only available when using whisper-1.

Word timestamps are insanely useful for large calls with interruptions (e.g. multi-party debate/Twitter spaces), allowing transcript lines to be further split post-transcription on semantic boundaries rather than crude VAD-detected silence. Without timestamps it’s near-impossible to make intelligible two paragraphs from Speaker 1 and Speaker 2 with both interrupting each other without aggressively partitioning source audio pre-transcription—which severely degrades transcript quality, increases hallucination frequency and still doesn’t get the same quality as word timestamps. :)

Even worse: proponents of the Neural Efficiency Hypothesis[1] might interpret the "mean brain activation" values reported in the study[2] in the exact opposite manner. :)

[1] https://en.wikipedia.org/wiki/Neural_efficiency_hypothesis

[2] https://jamanetwork.com/journals/jamanetworkopen/fullarticle...