I like "Proverb" as smaller than Haiku too, Aphorism is also good. But seriously I want Anthropic to up its small model game. Haiku is not competitive, Deepseek v4 flash outperforms my uses for about $0.10 / $0.20. Whereas Haiku 4.5 is $1/$5.

IMO Anthropic should just play the game at all the price tiers because it otherwise forces people to go elsewhere. I would probably pay for a "Proverb"/"Aphorism" class model that was worse than Deepseek at the same price just to stay in the ecosystem, if given the option.

(Note: I also see Google seem to make the same mistake, they actually do have competitive models in Gemma family but they don't make them available via the API. So there may be some reason for this.)

"Comeback" or "quip" for a low-latency sub-Haiku model.

What would you use Haiku for?

Models like Haiku are great if you just want to apply JSON structure to some free text.

For some reason I actually thought it was designed specifically to write Haikus!

Gemma 4 31b outperformed Gemini 3.1 Flash-Lite in our app benchmarks (agentic tool use via api in our application as a part of various workflows). But google won't let you pay to use Gemma models, you have to go elsewhere, I think this may be because it would cannabilize Flash-lite.

You can actually get the gemma-4 models on a per-token API basis, you just have to click some extra buttons (in GCP). Not the same for other open weight models. For those they make you run your own hardware.

Use OpenCode Go instead: https://opencode.ai/go

That doesn't have the Gemma models by the looks

They only host models they have evaluated and found good at coding

Curious logic. Does Google want you to use it or not? Do they want to be paid for tokens or not? why segregate open and closed?

It's not parameter size - there is apparently such a thing as "Gemini Nano", which famously is downloaded automatically by Chrome. How similar is it to Gemma E4B? And how strange - you have the weights, but you don't "have" them?

I think its even more puzzling because you can't even run Gemma 31b on google cloud, they only let you test it with a rate limit. No way (I can find) to actually pay them to use it.

We saw great results in our usecase using google direct. Moved to Openrouter because google wouldn't let us use it beyond a test.

Then Openrouters performance looked worse, not sure if there was a quantized version or something. So we instead looked at Deepseek v4 Flash, and opted to go for that.

This model would probably be great for a super low cost cloud model, would love to use it in the cloud, Google makes you go elsewhere.

I'm using it for one of my use cases (ocr) on openrouter right now.

It’s on openrouter. We just noticed performance was worse in a specific agentic app usecase. It’s possible we made an implementation mistake, my main point though is Google is really silly not hosting their own models.

I tested Gemma 4 31b for OCR and it's very good at it. This makes sense because I also get the best OCR results from Gemini compared to Claude or ChatGPT in my use case.

If we define product market fit as profitable with a trillion dollar valuation, I think the term has lost its helpfulness.

I do agree with the author that these companies seem much stronger financially recently though.

I can’t read the full article, but the snippet (6%gdp/$2T) seems not that expensive? And you could read that either way ”cheap so we should do it” or “if we end up needing to do it, we can do it”.

Cost is one thing, time is another. You’d need massive investment in trade education and then wait 20 years. See vocational education in German speaking countries as an example.

I am pretty convinced that friendly fraud is about 90% of chargebacks. I have seen some genuine fraud, but dwarfed by friendly fraud over time across 3 companies.

I think it does bear that out in general, although it is slightly more complicated. What seems to happen 1. Low-wage workers, as a collective group, experience an increase in earnings (Dube & Zipperer, 2024). 2. Total job losses do take place, but are minor and teens/part-time/new entrants workers lose more often (Belman & Wolfson, 2014; Redmond & McGuinness, 2024). 3. Lost hours & increased prices - businesses primarily absorb the cost by slightly reducing weekly hours worked & increasing prices for consumers (Redmond & McGuinness, 2024)

I would agree that modest minimum wage increases are far from the worst thing the government does, compared to other government interventions.

There are also home warranties or tech solutions/concierges like tidy.com.

The issue historically is that these concierge things are expensive (should be solved by tech/ai) and the warranties create their own class of problems (claim frustrations etc).

But home ownership is expensive, no way around it. But the work in coordinating etc doesn’t fundamentally need to be.

Yep, I debated mentioning home warranty services, but that is also a thing a lot of folks surprisingly don’t know about.

When I bought my house the realtor gave me a free home warranty for a year (a $600 value vs the tens of thousands in commission they got paid). The old washing machine broke during that year and the home warranty did replace it, so on paper I got good value out of the policy.

However, I think it was on the phone with the company waiting on hold for about four hours total to process the claim. That was without any dispute or confusion, it’s just how long the process took. The person on the other end of the line would ask me one question, then put me on hold for another 15 minutes before asking the next question. I got the impression they are actually handling multiple customers at once and switching between us. I can only imagine this system is designed to deter folks from making claims. I was afraid to delay the call any longer by inquiring about whether there were multiple replacement models I could pick between. I received a $600 model from a manufacturer I wouldn’t have chosen.

If I’d paid $600 for that home warranty, I would not be a satisfied customer. I certainly wouldn’t recommend home warranties as a time-saver. It would be much easier to directly purchase products and services from people with some incentive to obtain my business.

I definitely don’t think they’re generally worthwhile and don’t subscribe myself, but my in laws have had excellent luck with their policy.

The poor have dramatically more children than the rich on average, you don't need to be rich to have kids. Kids don't need to have rich parents to have a good live/upbringing.

I have long wondered why these companies have only one model? Why not many models, letting the user choose? Why not let users tune what they want to see? It seems like a better user experience and safer from this type of (IMO valid) concerns. I’d want nothing to do with picking a users feed if I were them.

Because better alignment with user interests would hurt the companies own interests. Their goal is to maximize profits, so even if they have multiple models available, they would choose the one predicted to maximize profits generated by you.