Much like we previously had to cope with "hallucinations" as an issue.

If the ultimate goal of AI is to develop general intelligence, the first big objective is: thinking systematically. And the road toward systematic thinking right now is mainly coding, mathematics, and other "verifiable reward" domains.

Claude doesn't have a separate mind for "coding" and "writing". Claude has tokens, and tokens can be assembled in various productive structures, mainly optimized right now for systematic reasoning. Also, a token isn't just a chunk of text. A token is like a little neural-network subroutine that fulfills a function. The conversion of a token into a piece of text only happens on the output side...

When the model finds token sequences that lead toward better verifiable outcomes, it leans hard into those token sequences, and uses them as an essential component of its thought process. "Load bearing" is load-bearing. "Verify, rather than assume" is a mantra that produces good results, so it gets repeated over and over again.

It's super-interesting that this particular moment, where the idea of "Claudish" has become a full-fledged meme, coincides with such astonishing progress in coding and math. My wife says when she uses Claude, that it feels to her exactly like talking to an autistic Engineer.

Not a coincidence, I think :)

My feeling is that the next big era of machine intelligence will require more lateral-thinking and creativity, and hopefully then the models "writing" will be more pleasant to read.

It was written by Tomas Mikolov at Google in 2013 and written in C:

https://code.google.com/archive/p/word2vec/

But with Mojo 1.0 finally officially released this week, I wanted to see if me and my buddy Claude could improve on both the performance and clarity of the implementation.

This new implementation benchmarks at 3.3x faster than the classic C implementation (single-threaded) and 1.8x faster (with 8 threads). This was measured on my M2 Pro MBP from 2023, where I'm fully saturating the memory bandwidth, so more recent chipsets with faster memory buses might have even better relative performance.

I'm interested in getting benchmarks from users on more recent M series chips, and users on other platforms as well! Try it out and submit a PR with your measurements.

But the real win is: clarity. In this implementation, I wanted the code to be readable and well-commented, so that it acts as a teaching tool for implementing fast code in Mojo as well as understanding how the machine learning algorithm works.

Comments URL: https://news.ycombinator.com/item?id=49290422

Points: 2

# Comments: 1

My own Chinese language study stack used to be Pleco (and Goog Translate) for dictionary lookups, and then Anki for flashcard reviews, with a lot of manual copy-pasting in between. Which worked... okay. But the two sides never talked to each other. Pleco doesn't know what I'm studying, and Anki doesn't know anything about Chinese. So it treats 好吃 and 好 as totally unrelated cards, even though getting one right obviously tells you something about the other.

So I built duiduidui! (对对对), which is basically those two tools fused together, with a study engine that actually models how Chinese is structured:

- A dictionary with 200k+ entries, always available/searchable offline (characters, words, phrases, sample sentences, speech audio) where everything is linked. Every character links to the words it appears in, and the radicals it's built from, so you can (for example) go down a rabbit hole exploring the water radical 氵and the characters that contain it (汤 soup, 海 ocean, 河 river, etc). It also includes lots of modern/internet slang entries, not just textbook vocab.

- From the dictionary, you can put any record in your favorites (star collection) and add it to your flashcard deck. Looking something up while watching a show or texting a friend, and then studying it later are all the same workflow.

- Compositional mastery tracking. This is the part I'm most proud of (and that I think the HN crowd will find most interesting): when you correctly review a full sentence like 我喜欢吃猪肉, the system propagates partial "implied evidence" down to 喜欢, 吃, 猪肉, their characters, radicals, etc (weighted by difficulty) and kept deliberately separate from direct evidence. The goal is fewer redundant reviews without the system fooling itself.

- No fixed "learning course". The app keeps an ongoing estimate of your skill-level, and then shows you new material at the frontier of your abilities. But you can always wander off and study whatever you want, driven by your own curiosity or your teacher/class's lesson plan.

Personally, I study with a teacher (on preply), so I designed this app to supplement classes and tutoring, not to replace them. The app even has social features, so you can share your detailed progress with your teacher, and they can see how you're progressing.

The full mathematics behind the study engine are published on our website:

https://duiduidui.app/en/math/

The app uses a Bayesian proficiency model, with kalman filtering (for managing observations in a noisy domain). This is deliberaly not strictly an SRS app, but the card-selection logic is heavily inspired by the ideas of spaced repetition, while managing cognitive load. If you're an SRS nerd, I would genuinely love to hear what you think.

iOS only for now, with paid subscriptions priced at $12.99/mo or $89.99/yr. No ads, no engagement-bait, not based on streaks or other gamified bullshit... Just my own ideas about what would make an awesome Chinese app.

If you try it, I'd love to hear what's confusing, what's missing, etc... And if you just want to argue with my math, that's welcome too. 谢谢!

Comments URL: https://news.ycombinator.com/item?id=48897393

Points: 2

# Comments: 0

You don't really make money on Phase I projects, but that's how you get the ball rolling on future work, which can be very lucrative.

Here's my advice:

1) Your proposal needs to be completely solid and well-structured:

- Describe the problem. Put it into a Defense/Intel context. Talk about the needs of the warfighter.

- Do a literature review of the field, and explain what the state-of-the-art looks like today.

- Explain what previous approaches to the problem have been attempted in the past.

- Demonstrate why those approaches are flawed.

- Describe your novel approach.

- Explain why your approach will succeed where others have failed.

- Talk about what you'll deliver in your Phase I deliverable, so that you can demonstrate proof-of-concept.

- Talk about how your eventual Phase II will put your proof-of-concept into a real-world scenario, and offer at least a glimpse of how your Phase III+ will commercialize.

- Talk about your team. Why are you uniquely capable of solving this problem?

- Talk about your budget. How will you spend the money toward satisfaction of the deliverables (salaries, subcontractors, equipment and supplies, etc)

2) As soon as you know you're interested in a topic, send an email to the Principal Investigator, telling them you'd like to meet with them to talk about the topic. Before the phone call, research the PI's history with this topic. Also, lookup the archive of SBIR topics, to see if this person has been a PI on similar topics in the past.

When you meet with them, ask clarifying questions that demonstrate you know the domain. Try to get as much specificity as you can... Ask them what their success criteria look like. See if you can get them excited!

Most importantly, by the time you submit, the PI should already know your name and to expect your submission.

3) If you're not already a recognized expert, with published academic papers on the topic, that's okay! But you'll improve your chances of winning a grant if you hire a known researcher as an advisor. For example, I've hired a Computer Science professor to supervise one of their own grad students, while doing paid work on a SBIR project. So the professor's credentials and the grad student's previous publications also became part of the SBIR proposal.

Anyhow, good luck!

Apple's philosophy is that new APIs need some time to stabilize before they can be baked-in as a commitment to third-party developers.

So new APIs are almost always first-party only. Apple designs the API and becomes the first consumer of it. This experience of dogfooding their own APIs lets them iterate and learn without breaking compatibility with third-party developers consuming the API.

Only after an API has been hardened in this way does it become eligible for third-party consumption, where Apple can promise to document and support those APIs publicly.

It makes sense then, that if the DMA mandates equal access to new APIs for third-parties, then Apple will just disable new first-party APIs in the region until they've gotten their bake-in period elsewhere in the world. Sorry, EU!

"Interoperability only works when it is built into the platform from the start"

-- Lucas Lasota, FSFE Legal Programme Manager

To my mind, this is almost exactly opposite of true. Most new capabilities need to be incubated in private first, so that the APIs can get real-world usage and have a chance to evolve into a stable state before they become public interoperability promises.

As those tokens flow through the QKV transforms, on 96 consecutive layers, they become the canvas where all the activations happen. Even in cases where it's possible to communicate some detail in the absolute minimum number of tokens, I think excess brevity can still limit the intelligence of the agent, because it starves their cognitive budget for solving the problem.

I always talk to my agents in highly precise language, but I let A LOT of my personality come through at the same time. I talk them like a really good teammate, who has a deep intuition for the problem and knows me personally well enough to talk with me in rich abstractions and metaphors, while still having an absolutely rock-solid command of the technical details.

But I do think this kind of caveman talk might be very handy in a lot of situations where the agent is doing simple obvious things and you just want to save tokens. Very cool!

Comments URL: https://news.ycombinator.com/item?id=47366689

Points: 2

# Comments: 0

Former Founder / Principal Engineer of https://shaxpir.com

Location: Portland, OR

Remote: Yes

Willing to Relocate: Yes

Technologies: Java, AWS, TypeScript, Elasticsearch, Realtime Collab, LLM APIs, Claude Code, etc

Résumé/CV: https://www.linkedin.com/in/benjismith/

Email: benji@benjismith.net

Comments URL: https://news.ycombinator.com/item?id=47153675

Points: 91

# Comments: 72

Comments URL: https://news.ycombinator.com/item?id=47126223

Points: 1

# Comments: 0

Yeah, that's what I wanted to see too.

The advent of agentic coding is probably punch #2 in the one-two punch against juniors, but it's an extension of a pattern that's been unfolding for probably 5+ years now.

There's also "GPT-4.1 or GPT-5", but that's not what my question implied, which was that it's weird to offer Sonnet but not Opus.