mchusma

9 ms·

151.
by mchusma 6mo ago
At first, I was more excited about the Flash model, but I'm now more excited about the Pro model in many ways. I feel like the Pro model with an Run through unsloth, and with some fine tuning, is gonna be enough for many vertical SaaS
152.
by mchusma 6mo ago
I honestly think health data should be public by default to any health researcher. We should do whatever we can to solve disease and live forever. Privacy be damned, I want life.
153.
by mchusma 6mo ago
For comparison on openrouter DeepSeek v4 Flash is slightly cheaper than Gemma 4 31b, more expensive than Gemma 4 26b, but it does support prompt caching, which means for some applications it will be the cheapest. Excited to see how it compa
154.
by mchusma 6mo ago
I appreciate this, makes me trust it more than benchmarks.
155.
by mchusma 6mo ago
Kind of? But I really care about price speed and quality. If it used 10x tokens at 1/10th the tokens and same latency I would be neutral on it. Kimmi 2.6 for example seems to throw more tokens to improve performance (for better or wors
156.
by mchusma 6mo ago
I’m pretty convinced that this is what Kimmi 2.6 does, mostly just thinks more.
157.
by mchusma 6mo ago
I like both GLM and Kimmi 2.6 but honestly for me they didn’t have quite the cost advantage that I would like partly because they use more tokens so they end up being maybe sonnet level intelligence at haiku level cost. Good but not quite a
158.
by mchusma 6mo ago
You can run models on FPGAs and get massive cost, speed, and throughput gains (like 10x). The reason people don’t do it is because of other improvements (algorithmic) means that nobody really thinks locking into a model makes sense…yet. Wou
159.
by mchusma 6mo ago
Closer to free than now…so I support. I would prefer us automakers to start competing, but at this point I’ve given up on Them.
160.
by mchusma 6mo ago
I think one issue I have with Zone 2 proponents is that it makes sense that if you can do a lot more stuff at Zone 2 than you can at less. For most endurance athletes, the volume of running beats everything else in terms of efficacy. Howeve
161.
by mchusma 6mo ago
Ever since I saw the first one of these one-bit models made by Microsoft, I thought this was a fascinating route. I assume that in practice, this is less helpful than it seems, just because there's every economic incentive in the world
162.
by mchusma 6mo ago
i think as the pricing has gone up on the Chinese models it has made them less appealing, and with the introduction of Gemma-4 not many are at the pareto frontier (also in my experience, not just the stats): https://arena.ai&#x2F
163.
by mchusma 6mo ago
it feels like between K2.6 and GLM5.1 we have Sonnet level intelligence at roughly Haiku level pricing. Which is great. I'm hoping that Anthropic will be able to release an updated Haiku soon and they really need something that is 1&#x
164.
by mchusma 6mo ago
You should do what you want, and as a break it’s fine. But IMO right now the most leverage for most people is learning how to effectively manage agents. It’s really hard. Not many are truly good with it. It will be relevant for a long time.
165.
by mchusma 6mo ago
Anthropic has been avoiding the hard thing, but they just need to do SOME kind of pricing thing here to shift demand. I expect there is just no amount of tricks that can handle the few hours at peak load. They need "surge pricing"
166.
by mchusma 6mo ago
Maybe they should not allow routines during peak for now? might help shift load.
167.
by mchusma 6mo ago
I do think the Iran crisis should continue to push countries towards nuclear + solar. Like Ukraine helped shift some in Europe back to supporting nuclear after foolishly shutting down reactors.
168.
by mchusma 6mo ago
developer jobs are up: postings ~11% YoY, tech workforce +1.9% projected for 2026, BLS 15% long-term growth. Rebound from 2024-25 lows Jevons paradox
169.
by mchusma 6mo ago
I agree this is helpful, and the numbers are better. I think this is one of those situations where everyone is a bit right: 1. Its a meaningful jump 2. Its not quite as extreme as some people will say. Opus might have gotten there with more
170.
by mchusma 6mo ago
I did a bit for fun and felt there were areas where I did do well. However, the problems with these markets is the contracts are often such where you can win or lose on obscure technicalities. I bet on 5 things, and 4 were resolved in ways
171.
by mchusma 6mo ago
Without knowing much this seems kind of reasonable? Some baseline of care but not enough that most people will opt it. If I could make universal healthcare in the US it would be something like: - free generic meds where there are at least 3
172.
by mchusma 6mo ago
I think teslas issue is that they need the AI5 chip for robotaxi ops, the current chip just doesn’t cut it. So if they have batches end of 2026 and start optimizing the models, by mid 2027 volume production you might have robotaxis coming o
173.
by mchusma 6mo ago
I think the way I and others use it is code with clause, review or bug hunt with codex. Then I pass the review back to Claude for implementation. Works well. Better than codex implementation and finds gaps versus using Claude to review itse
174.
by mchusma 6mo ago
When people talk about peptides, they typically mean either (1) GLP compounds like Reta that have successful phase 3 trial data and will be approved or (2) stuff like 157 which has no evidence or plausible mechanism to work.
175.
by mchusma 6mo ago
I’m expecting the new iPhone release this fall to be coupled with some great version of Siri/model. This could be the first reason I’ve seen to upgrade in a while (although even that I’m not sure of, as I am king of in the “always use
176.
by mchusma 6mo ago
Feels like it is a “supply chain risk” using the same logic Anthropic was labeled such. (Note I do not think any should be a supply chain risk, but if they were being consistent)
177.
by mchusma 6mo ago
Doing a bit more research, this looks like it might perform roughly as well on text tasks with modest context windows, so may be just a better cheaper option unless you need a million token window.
178.
by mchusma 6mo ago
For those curious, on openrouter this is $0.14 input and $0.40 output, or ballpark half of Gemini flash lite 3.1 (googles current cheapest current gen closed model)
179.
by mchusma 7mo ago
I believe the biggest benefit of going to space, particularly in building space stations, is making humanity focused on building a bigger pie. This is one step towards this. But once we can build (effectively) infinite land, we will be in t
180.
by mchusma 7mo ago
I was excited about the 1.58 bit models from a year or two ago, but the never seemed to go anywhere. Curious in particular how this scales up.

More ›