Des Traynor

🚨Huge day for @Fin_ai🚨 Our AI team have been quietly working on Apex for quite some time. It's extremely exciting to finally tell the world about it. More here 👇

@DamirWallener I don’t have the time to do formal DD. I’m a small cheque early investor I ask a few questions, and usually get a sniff of lies.

@destraynor Isn’t the real question here…why aren’t you doing due diligence? How can you have time to invest but not have time to “inspect”…?

Startups misreporting ARR is frustrating as an infrequent angel investor who relies on trust (no time to inspect) ~Every startup claiming 0→$[X]M ARR in [Y] months always forget to clarify that it's either a) not really Annual b) not really Recurring c) not really Revenue

@mutlu82 I hope so, but I usually am out by that stage

@destraynor Isn’t this coming up in DD?

@idhawal Letters of intent. Pilots. Pro serve.

@destraynor Curious - If it is not “Revenue” then what are they even reporting revenue as? Cost of tokens? Credits?

@ananbatra 💯it’s “the boy who cried ARR hyper growth” now.

Also hurts founders with real ARR!

@KeithMillsD7 Basically this

Winning the internet today.

@sandhya Great piece

@danobrien20 🎯

Is Irish society more polarised than usually believed? This is just a theory: many of those in the business economy feel alienated and unrepresented by government. They endure some of the highest marginal rates of personal tax in the world, which were never restored to pre-2008, while these taxes are spent by government in ways they feel they don't benefit and are sometimes wasted. At the other pole are the hundreds of thousands in the public sector and NGOs who have a woke or woke-adjacent world view. They are happy with bigger government, long ago restored pay, high taxes and a strong socially liberal agenda in policy. The are more insulated from the vagaries of the market economy. These two groups have little in common and don't interact much. Just a theory.

@Clearpreso + we thought we were allowed chat shit about people with zero repercussions.

The EU position on most stuff, until about a few months ago seems to be "we just thought things weren't going to change, and if they did, they'd change in a nice way"

@emilyross Basically 🎯

I did indeed. But I needed Grok to dumb it down for me. How'd it do? Big AI programs (like the ones that chat with you) are built from many layers. Inside each layer, there is a part called "attention." Attention helps the AI look back at all the words it has seen so far and decide which ones are important for the next word it will make. To do this quickly when the AI is answering you, it keeps a special list in its memory called the KV cache. "KV" stands for "key" and "value." These are numbers that remember important details about every word the AI has already read.Normally, the AI splits its attention into several separate "heads." Each head looks at the words in its own slightly different way. Because the heads are different, the AI needs to store a full separate KV list for every head. This uses a lot of computer memory, especially when the conversation gets long.The team at Fin AI found a new way called LRKV (Low-Rank Key-Value).Here's what they do:They make one big shared list that all the heads can use together. This shared list holds most of the important information. Then they add a very small extra list for each head. This small list only holds the tiny differences that make that head special. Because most of the information is shared, the total KV cache now needs only about half the memory (45-53% less than before).The math is set up so the computer can still do the exact same calculations as the old way. Nothing important is lost or changed. The AI works at the same speed.They tested this on AI models from small (128 million parameters) up to bigger ones (6.3 billion parameters). The new way:Uses much less memory for the KV cache. Trains faster (it reaches good results in 18-25% fewer steps). Often ends up with slightly better overall performance on tests like math problems, science questions, and coding tasks. In short: They found a smart way to share most of the memory between the heads while still letting each head keep its own small differences. This saves a lot of computer memory without making the AI worse or slower. That's useful for companies that run these big AIs because they can handle longer chats or run on cheaper hardware.

Our AI group published a novel finding from pre-training research In short: you can cut KV-cache memory in half by sharing most of the attention structure across heads + keeping small per-head differences—without hurting model quality or speed I'll explain this as best I can

Hope you found this interesting, full write up here on our AI Research blog fin.ai/research/low-r…

We found that sharing most of the structure between heads (since they're looking at the same thing), while allowing a small footprint for meaningful differences gets something just as good, as fast, but using 50% less memory Large implications for labs doing big training runs

We’re excited to share our new form of attention, Low Rank Key Value attention. This is a drop-in replacement to standard MHA that in our tests, reduces KV-cache by ~50%, with even lower test loss, across many scales of experiment.

@bhalligan @buccocapital Transparency and tooling are essential here

The folks in the middle of the org chart whose main function is "carrying context" are in a tricky spot in world where folks start adapting something appraoching "Dorsey mode" a la Block, Coinbase, Ramp, etc. Good article by @buccocapital on it.

LLMs decide what languages and libraries we use. They’ll also decide what products a business runs on. They'll prefer ones that work with them. Every business now needs a CLI. fin(.)ai/cli

That has been my experience, based on the questions people often ask. Often it’s literally “how do I get my team to learn Claude Code” and I realise they haven’t carried a difficult ask in a long time.

@destraynor vast majority will fail on number 3

AI tests your company's 1 Engagement: who's paying attention in their role, company, industry over the past 4 years 2 Ambition: who wants to win, who just wants to stay alive, who's happy to fade away slowly… 3 Leadership: who'll make brave decisions and say unpopular truths