voxgen

Even the official provider on OpenRouter seems to have this issue. Hope it's an easy fix for them.

Pay attention to what it forgets, and start telling it to proactively note down those things into nominated files while it works, e.g. indexing topics covered by each chapter/section/page of each paper/material. GPT-5.6 is really good at finding context again, if given the smallest hint where to start looking.

You can also ask it to read your past sessions, find places where it was wrong after compaction, and figure out a strategy to persist the most important summary information.

I have several Hermes threads that have each had >10 compactions with a 200k context limit, and with the right instructions on note-taking, they require at most a "double-check that against our past decision records" to be put back on track after a slip-up.

It's a solvable problem if you're willing to throw more tokens at it. Frontier models have gotten very good at cleaning up their own messes. You just need the right skills/loops, and to stick to models that consistently follow instructions (i.e. GPT-5.5/GPT-5.6-Sol).

There is a third case where the other party doesn't realize that the asker lacks the relevant experience to discern good LLM answers from bad answers for that topic.

Same solution as case one though - don't be afraid to say "Claude said X but that doesn't sound right".

Denying tech export to cooperative allies is certainly a move.

Many nations are now likely thinking: Why cooperate on international IP enforcement if we get lumped together with adversarial nations anyway?

As a current free-tier user, this price & transition plan seems completely reasonable. I'll upgrade when the nudge comes. I've got at least that much utility from it so far. Very glad this isn't a subscription.

I have found Claude to be especially unpredictable. I've mostly switched to GPT-5.4 now - although it's slightly less capable, it's massively more reliable.

Some of the flak is that issues are often only acknowledged once a fix is in place, and the partial fixes are presented as if they solve the whole problem.

The near-instant transition from "there is no problem" to "we already fixed the problem so stop complaining" is basically gaslighting. (Admittedly the second sentiment comes more from the community, but they get that attitude after taking the "we fixed all the problems" posts at face value.)

Thank you for the reminder.

Childhood, at least IME, does a bad job of preparing people for this: fault and blame rarely matter in real interactions. When no teacher is around to play judge, all that matters is that you get a favorable outcome. That's something you need to figure out, usually in the heat of the moment, how to get. It takes active effort to understand the other side.

It works for me Firefox's Cloudflare DNS over HTTP.

For clarity, the recent issue[0] likely wasn't intermittent. Cloudflare's malware blocking DNS server now blocks those archive.today sites. Doesn't affect the non-malware-blocking DNS server (1.1.1.1).

[0] https://news.ycombinator.com/item?id=47474255 "Cloudflare flags archive.today as \"C\&C\/Botnet\"; no longer resolves via 1.1.1.2"

This could be an explanation for the drama - LLMs are trained to learn and emulate correlations in text.

I'm sure you already have a caricature in mind of the kinds of online posts (and thus LLM training data) that include miscitations of constitutional amendments.

It's not perfect but it does have a few opt-in security features: running all tools in a docker container with minimal mounts, requiring approvals for exec commands, specifying tools on an agent by agent basis so that the web agent can't see files and the files agent can't see the web, etc.

That said, I still don't trust it and have it quarantined in a VPS. It's still surprisingly useful even though it doesn't have access to anything that I value. Tell it to do something and it'll find a way!

I'm working in AI, but I'd have made this anyway: Molty is my language learning accountability buddy. It crawls the web with a sandboxed subagent to find me interesting stuff to read in French and Japanese. It makes Anki flashcards for me. And it wraps it up by quizzing me on the day's reading in the evening.

All this is running on a cheap VPS, where the worst it has access to is the LLM and Discord API keys and AnkiWeb login.

Ratio/quantity is important, but quality is even more so.

In recent LLMs, filtered internet text is at the low end of the quality spectrum. The higher end is curated scientific papers, synthetic and rephrased text, RLHF conversations, reasoning CoTs, etc. English/Chinese/Python/JavaScript dominate here.

The issue is that when there's a difference in training data quality between languages, LLMs likely associate that difference with the languages if not explicitly compensated for.

IMO it would be far more impactful to generate and publish high-quality data for minority languages for current model trainers, than to train new models that are simply enriched with a higher percentage of low-quality internet scrapings for the languages.

It requires tax increases, and the average earner's UBI will typically balance out the tax increase, meaning they don't directly profit.

UBI isn't about giving everyone free money. It's about giving everyone a safety net, so that they can take bigger economic risks and aren't pushed into crime or bullshit work.

The upper half of society will only see the indirect benefits, like having greater employment/investment choices due to more entrepreneurialism.

That discussion also makes me worry that they may try to use LLMs or LLM-based metrics to measure the size of the gap as a proxy for value of the content.

The landlord of the marketplace should probably not dabble in the appraisal of products, whether for factuality or value.

without punishing regular browsing humans.

As a content consumer, I'm also hoping to be part of the ecosystem. I already use Patreon a lot as "AdBlock absolution", but it doesn't fix the market dynamics. Major content platforms tend to stagnate or worsen over time, because they prefer to sell impressions to advertisers than a good product to consumers.

What makes you think the secrets are small enough to fit inside people's heads, and aren't like a huge codebase of data scraping and filtering pipelines, or a DB of manual labels?

Please consider also describing the business model on the website, even if hidden away on a FAQ. I've so much subscription fatigue now, I just don't try things out if needing a subscription is an inevitability. I'm happy to pay for good products, just not happy to be forced to pay a fixed rate for continued access even if my usage dwindles.

If you are thinking of adding a one-off-donation-style purchase method, consider giving annual reminders to renew it. At least in my case, I'm not unwilling to pay repeatedly if development continues, just unwilling to make an upfront ongoing commitment.

I don't think retrofitting existing languages/ecosystems is necessarily a lost cause. Static enforcement requires rewrites, but runtime enforcement gets you most of the benefit at a much lower cost.

As long as all library code is compiled/run from source, a compiler/runtime can replace system calls with wrappers that check caller-specific permissions, and it can refuse to compile or insert runtime panics if the language's escape hatches would be used. It can be as safe as the language is safe, so long as you're ok with panics when the rules are broken.

It'd take some work to document and distribute capability profiles for libraries that don't care to support it, but a similar effort was proven possible with TypeScript.

The last major innovation as a product was PWA support starting in 2016.

Browsers used to try new ideas like RSS, widgets, shared and social browser sessions. Interfaces to facilitate low-friction integration with the rest of your life, and to multiplex data sources so that it's not a hassle to have many providers for [news, entertainment, social] experiences.

Likely no coincidence that this innovation languished once monopolies started pumping money into the ecosystem.

It's interesting that there are no reasoning models yet

This may be merely a naming distinction, leaving the name open for a future release based on their recent research such as coconut[1]. They did RL post-training, and when fed logic problems it appears to do significant amounts of step-by-step thinking[2]. It seems it just doesn't wrap it in <thinking> tags.

[1] https://arxiv.org/abs/2412.06769 "Training Large Language Models to Reason in a Continuous Latent Space" [2] https://www.youtube.com/watch?v=12lAM-xPvu8 (skip through this - it's recorded in real time)

Or is Behemoth just going through post-training that takes longer than post-training the distilled versions?

This is the likely main explanation. RL fine-tuning repeatedly switches between inference to generate and score responses, and training on those responses. In inference mode they can parallelize across responses, but each response is still generated one token at a time. Likely 5+ minutes per iteration if they're aiming for 10k+ CoTs like other reasoning models.

There's also likely an element of strategy involved. We've already seen OpenAI hold back releases to time them to undermine competitors' releases (see o3-mini's release date & pricing vs R1's). Meta probably wants to keep that option open.

My thoughts go out to the poor engineers who got put on call because someone scheduled a product release on the day before the biggest holiday of their year.

It's not even "nearly as good as o1". They only compared to the older 4o.

You can safely assume Qwen2.5-Max will score worse than all of the recent reasoning models (o1, DeepSeek-R1, Gemini 2.0 Flash Thinking).

It'll probably become a very strong model if/when they apply RL training for reasoning. However, all the successful recipes for this are closed source, so it may take some time. They could do SFT based on another model's reasoning chains in the meantime, though the DeepSeek-R1 technical report noted that it's not as good as RL training.

Vegetarian keto is certainly possible, but vegan would be very tough. Only 2 out of 6 of my regular meals[1] have meat in them, and I'd probably replace these with tofu and mushrooms if I could tolerate them. There's a world of keto&vege analogues to try for noodles, breads, and pizza bases. IMO some of them are nicer than the carby versions.

I also struggle with willpower and it took me ~10 big attempts over ~14 years before I managed to stay on it long enough to fix my metabolism. I just wanted to spread the message of hope that every attempt gets easier. Mindset plays a big role - I've seen a few people push themselves really hard then declare it impossible and never give it another shot. If you know you're playing a long game, take a break if you're really suffering, and don't beat yourself up over failures, it's easier to try again next time you have the energy.

[1] I've lots of intolerances - whitelisting was easier than blacklisting. Here's the list: flaxmeal porridge, keto bread + cream/cottage cheese, omelette w/ pizza toppings, egg & cheese salad, caesar salad (w/ chicken), mince+vege+cheese mealprep'd casserole

I'm at 3 years with occasional breaks. At a certain point my weight wouldn't go lower and I started feeling terrible. I think I was producing more ketones than I could use. I'm not sure exactly what fixed it, but now I'm sustaining a low-carb, low-but-nonzero-ketone mode, and still getting 50-75% of the mental/energy/anti-inflammatory advantages.

I think it was either changing my diet to focus on veges instead of meats (still 15-30g net carbs/day though), or adding artificial sweetener to maybe fool my body to making insulin? The science says that shouldn't happen, but idk what else it could be.

Don't give up! Induction gets easier every time, and you learn lots of tricks/recipes, like keto-ade to feel better during induction, and making oats/flaxmeal tasty for cheap & quick breakfasts. You don't have to commit to long streaks, or feel bad about sunk cost when you cheat. All that progress accumulates.

I've been in and out so often now I can happily switch between keto at home & unrestricted on vacation/occasions. At worst I get 1 day of dopiness starting carbs, and 1 day of mild cravings stopping them, but usually I don't even notice.

it's not clear if that was the author's actual intention

The paper[1] doesn't appear to have any other connections to the book/response/memes. A clear distinction is that the UB paper very directly and prominently states the question, rather than cloaking it in allusion or having a lengthy preface trying to contextualize it.

[1] https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2024/p34...

Her channel has strayed far beyond the topics she has credibility in.

I appreciate that she makes her videos so easily verifiable, by prominently showing her research, that it was easy to see the point when this started happening and tune out. A lot of opinion-faucets on the internet try to be irrefutable by hiding their sources.

I don't trust Sabine intrinsically, but I trust that I can notice when she under-researches a topic or makes a leap of logic. She conveys enough good information that I find it worth my time to watch.