tempusalaria

Most of EA’s revenue comes from franchise games that are way below typical AAA standard. EA’s value is from IP not talent

Lots of situations, here are 2 I’ve faced recently (cannot give too much detail for privacy reasons, but should be clear enough)

1) low latency desired, long user prompt 2) function runs many parallel requests, but is not fired with common prefix very often. OpenAI was very inconsistent about properly caching the prefix for use across all requests, but with Anthropic it’s very easy to pre-fire

All these things are designed to create lock in for companies. They don’t really fundamentally add to the functionality of LLMs. Devs should focus on working directly with model generate apis and not using all the decoration.

I vastly prefer the manual caching. There are several aspects of automatic caching that are suboptimal, with only moderately less developer burden. I don’t use Anthropic much but I wish the others had manual cache options

A lot of the current code and science capabilities do not come from NTP training.

Indeed in seems in most language model RL there is not even process supervision, so a long way from NTP

Cerebras has very limited scale. Mistral has very few users so they can use cerebra’s in inference whereas OpenAI and Anthropic cannot. If mistral grows a lot they will stop using cerebras

Fast tire changes only matter a very limited amount of the time (pretty much only if the extra time drops you a place, so there has to be 1 car/20 in a specific 1 second window on what is typically a 90s lap for 3s (a slow stop) vs 2s (a fast stop) to matter. Maybe 20% of the time a slow stop happens, it costs a driver.

Strategy matters a lot and good strategy is worth at least a few positions in a race.

I imagine it runs civ 2 pretty well

WhatsApp is certainly worth less today than what they paid for it plus the extra funding it has required over time. Let alone producing anything close to ROI. Has lost them more money than the metaverse stuff.

Insta was a huge hit for sure but since then Meta Capital allocation has been a disaster including a lot of badly timed buybacks

SFT is part of the classic RLHF process though

Yes this write-up is not about agents.

In fact it’s a great illustration of why the hype around agents is misplaced!

I understand that calling it ‘agentic’ is nice for marketing, but most of what is described in this blog post is not related to agents. The design patterns you describe are explicitly non-agentic. Many of the use cases described are better handled by a single LLM call rather than an agent.

Finally, saying that agents can have predictable behavior is wrong (except on simple tasks where you shouldn’t be using an agent anyway). Agents loop and compound their input, making them highly non-deterministic even for the same prompt.

They may not be acting in good faith but there is extremely clear evidence that UCLA has engaged in illegal racial hiring and admissions practices and has supported antisemitism on campus. UCLA chose to give them that ammunition.

if you are p testing this isn’t the case. A positive result is a much stronger assertion

The term agent is just way overloaded. This guy defines it completely differently the the big labs, and I’ve seen half a dozen different definitions in the last few months.

In the long run the definition used by OpenAI, Anthropic et al will win out so can we just all switch to that?

Even as someone who is skeptical about LLMs, I’m not sure how anyone can look at what was achieved in AlphaGo and not at least consider the possibility that NNs could be superhuman in basically every domain at some point

I agree I find claude easily the best model, at least for programming which is the only thing I use LLMs for

SemiAnalysis has made up many things.

They claim that a small Chinese hedge fund could acquire $1bln in GPUs, with no state support, including many sanctioned chips, then trained a model optimized for a far smaller server compute size, and that they have a source at this very small fund who is willing to admit to export violations. A 40bln param active model is exactly the size you would expect from a server of the size they claim.

What’s more likely - that semianalysis made it up like they have a bunch of other things, or that all the above is true?

SemiAnalysis is wrong. They just made their numbers up (among many other things they have invented - they are not to be trusted). I have observed many errors of understanding, analysis and calculation in their writing.

Deep Seek R1 is literally an open weight model. It has <40bln active parameters. We know that for a fact. That size of model is definitely roughly optimally trained over the time period and server times claimed. In fact, the 70bln parameter Llama 3 model used almost exactly the same compute as the DeepSeek V3/R1 claims (which makes sense, as you would expect a bit less efficiency for the H800 and for the complex DeepSeek MoE architecture).

DeepSeek v3 (where the training cost claims come from) was announced a month ago and it had no impact outside of a small circle

Texas is a world leader in renewable energy. Easy permitting, lots of space, lots of existing grid infrastructure from the o&g industry.

1) DPO did exclude some practical aspects of the RLHF method, e.g. pretraining gradients.

2) the theoretical arguments of DPO equivalence make some assumptions that don’t necessarily apply in practice

3) RLHF gives you a reusable reward model, which has practical uses and advantages. DPO doesn’t have useful intermediate product.

4) DPO works off preference, whereas desirable RL objectives could have many forms

in practice big labs are testing all these methods to see what works best.

The reality is that these are not culturally significant institutions and most people in London don’t care. Ordinary Londoners rarely use these markets, and they mostly sell to restaurants, and require major financial support.

Not everything from the past is good. We don’t still run coal power plants in London just to get that historic smog.

They were already planning to move both markets way out of central London (to a place no-one but a local would ever go) this is just changing it to not providing a new site.

many of these labs have more funding in theory than OpenAI. FAIR, GDM, Qwen all are subsidiaries of companies with $10s of billions in annual profits.

Airbus was a company setup by consolidating companies controlled by some of the most powerful countries in the world, which sold planes to captive state airlines and militaries controlled by those same governments and their allies.

What an insane comparison.

This is very similar to how LLMs are taught to understand images in llava style models (the image embeddings are encoded into the existing language token stream)

Only if it is reliably correct.

Google does offer an AI summary for factual searches and I ignore it as it often hallucinates. Perplexity has the same problem. OpenAI would need to solve that for this to be truly useful

Definitely they will.

OpenAI’s potential issue is that if Google offers tokens at a 10% gross margin, OpenAI won’t be able to offer api tokens at a positive gross margin at all. Their only chance really is building a big subscription business. No way they can compete with a hyperscaler on api cost long run

I also like 3.5 sonnet as the best model (best ui too) and it’s the one I ask questions to

We use Gemini flash in prod. The latency and cost is just unbeatable - our product uses llms for lots of simple tasks so we don’t need a frontier model.

It’s not clear this is true because reported numbers don’t disaggregate paid subscription revenue (certainly massively GP positive) vs free usage (certainly negative) vs API revenue (probably GP negative).

Most of their revenue is the subscription stuff, which makes it highly likely they lose money per token on the api (not surprising as they are are in price war with Google et al)

If you have an enterprise ChatGPT sub you have to consume around 5mln tokens a month to match the cost of using the api on GPT4o. At 100 words per minute that’s 35 days on continuous typing which shows how ridiculous the costs of api vs subscription are.