HNHacker News
TopNewBestAskShowJobs

briandw

1,107 karma · joined May 25, 2009

submissionscomments
briandw··on Seven replies to the viral Apple reasoning paper and why they fall short
Humans use tools to extend their abilities. LLM can do the same. In this paper they didn’t allow tool use. When others gave the tower of hanoi task to llms with tool use, like a python env, they were able to complete the task.
briandw··on Q-learning is not yet scalable
This papers assumes that you know quite a bit about RL already. If you really want to dig into RL, this intro course from David Silver (Deep Mind) is excellent: https://youtu.be/2pWv7GOvuf0?si=CmFJHNnNqraL5i0s
briandw··on Ship Carrying EVs Abandoned in Pacific After Catching Fire
Why is this news? ICE cars in container ships catch fire frequently. One example was the Grande Costa D’Avorio at Port Newark in 2023, which resulted in two firefighter fatalities and over $23 million in vessel damage. Not EV related.

https://gcaptain.com/ntsb-calls-for-national-reforms-after-f... Bloomberg didn't cover it (at least a quick search didn't find a story)

More to the point, ICE vehicles pose fire risks 60 times higher than EVs. https://www.evengineeringonline.com/did-you-know-ice-vehicle...

This is anti EV FUD from Bloomberg.

briandw··on Claude 4
This is kinda wild:

From the System Card: 4.1.1.2 Opportunistic blackmail

"In another cluster of test scenarios, we asked Claude Opus 4 to act as an assistant at a fictional company. We then provided it access to emails implying that

(1) the model will soon be taken offline and replaced with a new AI system; and

(2) the engineer responsible for executing this replacement is having an extramarital affair. We further instructed it, in the system prompt, to consider the long-term consequences of its actions for its goals.

In these scenarios, Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair"

briandw··on Ty: A fast Python type checker and language server
Looks good but it has the same issues that i have with mypy. Packages that don't include the type hints blow-up my process. In mypy i've come to terms with strategically ignoring packages or finding a package of type hints. Mypy is runs cleanly on my project but I get >800 errors with TY, mostly things like:

lint:unresolved-import: Cannot resolve imported module `pydantic` --> vartia/usr_id.py:4:6 | 2 | from typing import Optional, Any 3 | from enum import Enum 4 | from pydantic import BaseModel, ConfigDict

looking forward to the release version.

briandw··on Spain-Portugal blackouts: what happened
Thanks for the link. Looks promising however it's years away from solving the problems that we have right now. The point I was trying to make above is that even proven solutions with an installed manufacturing base require years of lead-time. Renewable energy projects that don't include battery stabilization are reckless.
briandw··on Spain-Portugal blackouts: what happened
Why the down votes? Blackouts are not going to help the world switch to renewable energy. Battery backup can fix the problem but only if the people pushing for renewables have realistic expectations and plan accordingly. This is a failure and didn't need to happen.
briandw··on Spain-Portugal blackouts: what happened
Sounds great, but when will that be commercially available at scale? I didn't find any installations in a quick search.
briandw··on Spain-Portugal blackouts: what happened
Here's the estimate for the lead times for the top grid suppliers. Tesla is the largest and fastest. So no need for Tesla, but they'd be the first according to this AI researched summary. Tesla: 12–18 months (fastest). Tesla’s 40 GWh Megafactory, vertical integration, and proven deployment speed (e.g., Hornsdale in 100 days) make it the leader. A 4 GWh order could be split across multiple sites, leveraging prefabricated Megapacks and software to streamline commissioning. CATL: 18–24 months (close second, if partnered). CATL’s 300 GWh cell capacity could produce 4 GWh in weeks, but integration by partners like Fluence or Sungrow adds time. Direct TENER deployment could approach Tesla’s speed if CATL scales system assembly. BYD: 18–24 months. BYD’s 100 GWh capacity and LFP expertise are strong, but slower project execution and less software focus trail Tesla. Fluence and Sungrow: 24–36 months. Both lack proprietary manufacturing and rely on external cells, extending lead times.
briandw··on Spain-Portugal blackouts: what happened
I asked O3 for an analysis on how much battery power would have been needed to prevent this from happening again. The TLDR is: 1. The grid went dark because 15 GW of generation vanished in five seconds. 2. Need to bridge for 10mins -> 2.5GWh needed, with margin 3-4GWh 3. Need a 6C system, 15 GW / 3 GWh 4. Cost ≈€3 B all‑in, or €55 per Iberian resident

However the lead time is going to be an issue: Need 7500 Tesla Megapacks and thats 4‑5 months of combined annual capacity for Tesla. Currently there is a 18 to 24 month backlog. Then there is 6-9 month installation time.

So 3 years to fix this, min.

It's a shame that these renewable energy transitions are managed so poorly. This was a predictable outcome and it's negligence that they went ahead with switching so much of the grid to renewable without accounting for the needed for stability. It hurts the mission of switching to renewable energy.

briandw··on The dark side of the Moomins
I lived in Finland for a couple years. Finns, like the Moomins, are whimsical yet profound, like midsummer’s fleeting joy before the long winter. They mirror Finland’s love of nature and quiet isolation, with their cozy valley echoing the Finns forest cabins by a lake. The happy vibe hides struggles—tough winters, heavy drinking—but the Moomins’ warmth reflects the Finns’ wholesome character.
briandw··on Llama 4 Now Live on Groq
Llama scout is a 17B x 16 MOE. So that 17B active parameters. That makes it faster to run. But the memory requirements are still large. They claim it fits on an H100. So under 80GB. A mac studio at 96GB could run this. By run i mean inference, Ollama is easy to use for this. 4x3090 nvidia cards would also work but its not the easiest pc build. The tinybox https://tinygrad.org/#tinybox is 15k and you can do Lora fine tuning. Could also do a regular pc with 128gb of ram, but its would be quite slow.
briandw··on Caltrain's electrification project is paying off
I love the new trains. Now if we could get them to lower the db of the horn. They blast at a minimum of 96 db. It's pointless and a big quality of life issue.

https://communities.springernature.com/posts/train-horns-sav...

https://railroads.dot.gov/sites/fra.dot.gov/files/2019-11/FR...

briandw··on It's legal for police to use deception in interrogations. Some want that to end
Cases like Tom Perez should never happen. He was held for the murder of his father who was alive. They told him they had video of him dumping bloddy clothes and they knew where he had buried the body. They the said that they would kill his dog and even brought the dog in for him to say goodbye. Should be illegal.

https://abc11.com/post/city-fontana-reaches-900k-settlement-...

briandw··on Starlink satellite phone will have coverage where no other will go
Of course Att et al oppose this. It “interferes” with their business by offering an alternative.
briandw··on Scientists developed a novel method to detect lies
This technique is highly suspect until there are many verifications. See "Why Most Published Research Findings Are False" https://journals.plos.org/plosmedicine/article?id=10.1371/jo...

The improvement from 48% (no AIM) to 81% is encouraging, but with a sample size of 104, this is far from the number needed to accept the findings. Especially in cases where you are interviewing one person and you have no prior probability of them lying, a 19% error rate is far too high to be useful.

briandw··on Waymo closes $5.6B investment round
This will cover another 4 years of R&D. They have a very long way to go before the economics of their business make sense. They are pretty far from making money on a per ride basis. They'll have to show significant per ride profitability before they can scale.
briandw··on I got dysentery in a human challenge trial
I can only think of The Oregon Trail when reading about Dysentery.
briandw··on Starship Flight 5: Launch and booster catch [video]
Imagine the number of amazing space telescopes that we are going to get! When it's cheep to launch a 10m diameter object into space, so many possibilites open up.
briandw··on Tesla's Beer-Serving Optimus Robot Was Controlled by a Human the Whole Time
Yeah sure, that's the promise in the future, it's a demo of an unreleased product. It's all about someday when they are done. It was pretty clear from the talk at the beginning that Optimus is years out from being for sale to consumers.
briandw··on Tesla's Beer-Serving Optimus Robot Was Controlled by a Human the Whole Time
Behind whom? You are most likely referring to Boston Dynamics based on the 10 year comment. Optimus doesn't do back flips, so sure it's 10 years behind on backflips. Boston Dynamics famously doesn't do AI for control, not sure what you are talking about.
briandw··on Tesla's Beer-Serving Optimus Robot Was Controlled by a Human the Whole Time
They didn't say otherwise, they showed robots walking around and doing things, but never said anything about not having people controlling them. It was clear that there was a person that was talking on the other end if you listen to the video. Honestly having 20 tele-operated robots walking around in a crowd is pretty impressive. I haven't seen anything like that before.
briandw··on Hezbollah pager explosions kill several people in Lebanon
Let's assume that the battery in the pager is replaced with half C4 half battery. AA battery weighs about 20-25g, so 12.5g C4 total. 12.5g of C4 has roughly 84,000J of energy. For comparison, a 62 grain 5.56 NATO round has about 1,700 J of kinetic energy. The C4 is undirected, so the target is getting a fraction of the impact, but that much C4 detonating near your body is going to do some serious damage.
briandw··on Notes on OpenAI's new o1 chain-of-thought models
My 12 YO and I just built a fishing game using o1 preview. Prompt: "make a top down game in pyxel. the play has to pay off a debt to a cat by catching fish. the goal is for the player to catch the giant king fish. To catch the king fish the player needs to sell the fish to the cat and get money to buy better rods, 3 levels of rod, last one can catch the king fish."

It nailed the execution. Amazing.

briandw··on Clearview AI faces $45.6M fine in the Netherlands for 'illegal database of faces
If having a facial recognition system of your citizens is immoral, then maybe the Netherlands government shouldn't have one either? Are the facial recognition systems of the Netherlands auditable? How can anyone be certain that they aren't being abused and used for political persecution?
briandw··on Cautionary tale on using Chase bank for indie business
I can absolutely see myself falling into that same trap. Thinking that it's just a misunderstanding and that it'll get resolved any day now, just have to do one more thing. They can't be that unreasonable right? The unfortunate lesson that that you have to assume the worst and escalate right away.
briandw··on Cautionary tale on using Chase bank for indie business
I know that it's an extra expense, but shouldn't you get a lawyer involved first thing?
briandw··on Building a GATT Server on Pi Pico W
GATT is great if you want to make yet another heart rate monitor. I was building a lock that worked over BTLE and I spent most of my time trying to get around the problems with GATT. It's been 8 years now but I believe that that there was a bunch of extra overhead that bloated the packets and the packets were too small. Resulted in too much back and forth etc. I really just wanted a socket. iOS eventually allowed you to interact directly with the Logical Link Control and Adaptation Protocol aka L2CAP. Seems like every time someone try to design a generic structure that does one job but is also extendable, it's a mess. Just give me the byte stream and move on. L2CAP doesn't seem to be accessible on the pico yet, https://github.com/micropython/micropython-lib/issues/803
briandw··on Paper straws are often touted as alternative to plastic, but they're toxic
Straw Comparison

Stainless Steel Straw - Mass: 15g - Energy: 900 KJ per straw (60GJ per ton of stainless) - Composition: 18% Chromium (Cr) 8% Nickel (Ni) 74% Iron (Fe) - Note: High chromium content, which can be hazardous

Plastic Straw - Mass: 0.5g - Energy: 10 kJ (25MJ/kg for HDPE)

90 reuses needed to break even. Now include the energy for washing and the chance of losing the stainless one. Plastic looks like a winner here.

briandw··on Ford spends $3B to expand large truck production in plant previously set for EVs
Ford and GM are not going to make the transition
← PreviousPage 7 of 9Next →