alansaber

4 ms·

1.
by alansaber 5d ago
As anthropic/openai subscription allocations get squeezed you'll see more people using "second rate" closed models like grok. The token allowance with a Cursor subscription is crazy.
2.
by alansaber 5d ago
This is why I like books. No attempt at attention hijacking. Wheras every other modern app tries to claw your eyes out.
3.
by alansaber 5d ago
Yes but on the sliding scale of how deterministic is this task- writing, not very. coding, can execute and write some tests (even if gamed/bad).
4.
by alansaber 5d ago
This is a neat workflow. Incentivises you to proactively think about your writing, even as you cut it.
5.
by alansaber 5d ago
Well said, the issue being, the form factor of many AI products disincentivises this behaviour. To give an example, using a CLI tool with a frontier model to write a blog (actually fairly common), you have to re-open it in another editor, p
6.
by alansaber 5d ago
Fundamentally they just want the best effort>output ratio. You get information about this during exploration (hence "bias to action"), but you can also get information from taking a moment to step back. It's almost as if t
7.
by alansaber 5d ago
Sure. But it was a bitch to get access to a good maths textbook- either you forked out a lot of money, got a library loan (we had to renew ours every 7 days), etc. With an LLM there's good breadth of knowledge so if can quickly hop acr
8.
by alansaber 5d ago
looks like a nifty information retrieval approach. cool. temporal/versioned chronology is absolutely useful.
9.
by alansaber 6d ago
They were always going to do surveillance. At least now you get a trendy e-commerce site recommendation too.
10.
by alansaber 6d ago
Need to scrape the logs from some leading researchers first
11.
by alansaber 6d ago
"Your limb replacement technology is very good, but I believe humanity should transform into silicon-based beings, and so for that reason i'm out"- dragons den 2030
12.
by alansaber 6d ago
They're for-profit corporations above all.
13.
by alansaber 6d ago
"Slightly different"- no, a completely different realisation of intelligence to humans. LLMs can definitely do cool stuff though.
14.
by alansaber 7d ago
Ah yes, the peak of technology.
15.
by alansaber 8d ago
I sympathise completely but a ~3B parameter model and ~3T parameter model are going to exhibit very different behaviour, one can only infer so much large model behaviour from the former.
16.
by alansaber 8d ago
Geopolitics is very complex, but at least those bigger states have a lot to lose. Under mass proliferation I can see a small country with an ideological nemisis deciding to just nuke them.
17.
by alansaber 8d ago
We have a romantic view of intelligence, perhaps stemming from intuition within the context of scientific discovery. Given enough intelligence, and enough context, a brilliant person can have a stroke of inspiration that allows them to make
18.
by alansaber 8d ago
Machines have a much higher capacity for work than human beings. Saying that these proofs did not require equivelant intelligence, but benefitted from sheer volume, does not strike me as unreasonable.
19.
by alansaber 8d ago
"Why yes, we had to exfiltrate 100% of your data so we could vectorise it and improve recall by -0.3%"
20.
by alansaber 8d ago
We did something similar for document version control. The fundamental premise is the same: give collaborators the full context involved in (agentic or otherwise) decision making, rather than inferring it from document snapshots. https:&#x
21.
by alansaber 8d ago
Exactly, this is why we see such minimalistic memory systems being used. Information retrieval is bad, but improving information availability by pre-computing information retrieval is not the obvious fix people think it is.
22.
by alansaber 8d ago
Spiderman pointing
23.
by alansaber 8d ago
Exactly- without constraining the input, you're guaranteed to get slop. First-draft AI copy is usually 95% to-cut, 4% to re-sequence, 1% to keep. Use it to think about iterating, not sweep the solution space blindly.
24.
by alansaber 8d ago
Agreed. We've basically phased-out the phrase "RAG" for "harness" even though a lot of the discussion is still the data integration rather than agent behaviour (sometimes an MCP or some kind of graph). The argument
25.
by alansaber 8d ago
LLMs bleed context from the conversation/attachments into the output. This has always been a problem with no real solution other than some crafty iteration/loops.
26.
by alansaber 8d ago
Honestly it looks like the biggest unlock at this point will be court reform, not improvements from individual practitioners.
27.
by alansaber 8d ago
Watermarking is just one step. We could easily see regulatory crackdown, and the requirement to provide IE a whitelisted email to use a legal product.
28.
by alansaber 8d ago
Lawyers are client-facing. Client wants lawyers to use AI. There are justifiable reasons to use AI (in measured doses). Hence, firms will use AI. They already have all kinds of strategems to increase their billing which AI slots into.
29.
by alansaber 8d ago
Yes, there's been a massive explosion in pro-se litigation. Courts will adjust accordingly, many already have.
30.
by alansaber 8d ago
Basically, they can't afford to not compete. They also don't have to compete so hard, because they have the brand recognition.

More ›