DenisM

9 ms·

121.
by DenisM 10mo ago
You’re welcome. I don’t know if benchmarks, sorry.
122.
by DenisM 10mo ago
Multi agent collaboration is quite likely the future. All agents have blind spots, collaboration is how they are offset. You may want to study [1] - this is the latest thinking on agent collaboration from Google. [1] https://www.
123.
by DenisM 10mo ago
I hear you - model development might overcome the shortcomings one day. However the "waiting out" strategy needs a timeout. It might happen that agentic crutches around LLMs will bear fruit much sooner than high-quality LLMs arr
124.
by DenisM 10mo ago
Well, some things fade out and some do not. How do we decide which one it is? The reason I ask is that the pace of new things arriving is overwhelming, hence I was tempted to just ignore it. Not because things had signs of transience, but b
125.
by DenisM 10mo ago
Why do you think they will fade out?
126.
by DenisM 10mo ago
> But the user just wants answer; they'd not like; but alignment. And there it is - the root of the problem. For whatever reason the model is very keen to produce an answer that “they” will like. This desire to produce is intrinsic
127.
by DenisM 10mo ago
It is in their training set by now.
128.
by DenisM 10mo ago
Gibberish can be the model using contextual embeddings. These are not supposed to Make sense. Or it could be trying to develop its own language to avoid detection. The deception part is spooky too. It’s probably learning that from dystopian
129.
by DenisM 10mo ago
I keep asking ChatGPT to read and summarize HN front page while driving, and it keeps blundering. I don’t know if there’s a business for you in this, but I would pay. Of course I always have questions about the subject, so it become the who
130.
by DenisM 10mo ago
Is there a voice chat mode in any chat app that is not heavily degraded in reasoning? I’m ok waiting for a response for 10-60 seconds if needed. That way I can deep dive subjects while driving. I’m ok paying money for it, so maybe someone c
131.
by DenisM 10mo ago
There is no mention of llm there?
132.
by DenisM 10mo ago
It’s not innate. Purpose trained llm can be quite stubborn and not very polite.
133.
by DenisM 10mo ago
As compute becomes cheaper your argument becomes more and more true. But it only works if workloads remain fixed. If workloads grow at similar rates you’re back to the same problem.
134.
by DenisM 10mo ago
Scale-up solves a lot of problems for stable workloads. But elasticity is poor, so you either live with overprovisinoed capacity (multiples, not percentages) or fail under spiky load which often time is the most valuable moment (viral traff
135.
by DenisM 10mo ago
Continuous eval is unavoidable even absent model changes. Agents are keeping memories, tools evolve over time, external data changes, new exploits are being deployed, partner agents do get upgraded. Theres too much entropy in the system. C
136.
by DenisM 10mo ago
Good point. There will be a mixture of those who succeeded accidentally, or by taking calculated risks, Or by virtue of sill holding their shot together. Drawin will chew up the rest, which might be the majority.
137.
Agentic AI at Scale: Redefining Management for a Superhuman Workforce(sloanreview.mit.edu)
2 points by DenisM 10mo ago|0 comments
138.
by DenisM 10mo ago
> There is no such thing as an AI strategy. There is only Business Process Optimization (BPO). Here’s your Ai strategy: every few months re-evaluate agent fitness and start switching over. Remember backstops and canaries. Details: Busine
139.
by DenisM 11mo ago
No one exposes SQL to clients though. I think where Gql differs from sql is it’s at a higher level. SQL bleeds performance and data layout (e.g. normalizing, limits), GraphQL does not. It’s not clear if it’s high enough to abstract knowledg
140.
by DenisM 11mo ago
That might be an interesting extension to a dev environment or git - convert terse rust into semi-verbose explanation. Sort of like training wheels, eventually you stop using it.
141.
by DenisM 11mo ago
I’m reminded of classical LRU cache implementation - double linked list and a hash map that points to the list elements. It is a queue if we squint really hard, but it allows random access and reordering. Do we have durable structures of th
142.
by DenisM 11mo ago
It would be more productive for camera manufacturers to embed a per-device digital signature. Those care to prove their image is genuine could publish both pre and post processed images for transparency.
143.
by DenisM 11mo ago
Sorry, I was being lazy and on a phone. I checked now - it’s closer to 25%. Low crime is anecdotal from my personal connections.
144.
by DenisM 11mo ago
Malicious or erroneous actor can also drop your s3 buckets. Account change has stricter permissions. The key problem is that data loss is really bad pr which cannot be reversed. Overcharge can be reversed. In a twisted way it might even str
145.
by DenisM 11mo ago
They could have given us a choice though. Sign in blood that you want to be shut off in case of over spend.
146.
by DenisM 11mo ago
I heard Spain has 20% unemployment among the young and the violence problem did not happen. Didn’t check it though.
147.
by DenisM 11mo ago
if you haven’t done that yet look into “the green revolution”. The practice of blasting things with radiation is rather old. Some of the Most used crops are the product of that process, and yet are perfectly “organic”.
148.
by DenisM 11mo ago
You may be looking for a “commercial LCD display”.
149.
by DenisM 11mo ago
Being part of the model is not the same as controlling for that variable, or is it?
150.
by DenisM 11mo ago
> For men in NAS, higher baseline optimism levels were similarly related to longer life span (Table 1, NAS; P trend = 0.002). After adjusting for demographics, baseline health conditions, and depression, compared to the least optimistic

More ›