mmmBacon

I run a team. So for me this is like having 2-3 extra people to just do some exploratory work and prototyping. Having actual people to do this is a luxury I don’t have and can’t justify currently.

I can’t just copy-paste what comes out. But I have to say I’m able to get substantially more done as it saves me a lot of grunt work. You do have to learn how to use it like any other tool. I have found that it’s helped me sharpen a lot of my own skills in multiple areas and improved my understanding of systems I’m working on. I am able to learn new things much faster because it has a real-time feedback mechanism.

I can’t speak to your direct report but I’d be concerned about falling behind as a leader if you aren’t using it in the course of your regular work.

I disagree. Couldn’t do what I’m doing with GPT-4 but I can with 5.

I am personally paying $200/month for the ChatGPT pro version. It’s the best investment I’ve made in my career in a long time. It’s like having 2-3 direct reports. The alternatives provided by my companies are too slow and aren’t anywhere as good. If people start using it anywhere near how I’m using it, there will be profound changes to the workplace in the coming 3 years.

The Surfliner in Southern California is a great deal. Take the business class ticket. OC to San Diego is a breeze and you can relax and enjoy all beach towns with a glass of wine. Took it from OC to Santa Barbara. It takes about an hour longer than driving but so worth it. Can’t say enough good things about it. Views in all directions totally amazing. Trains are packed at times but it’s good to see.

I think high speed cross country rail travel is not very practical either to build or for travelers. The distances are too large, terrain too difficult and population densities are too low west of the Mississippi. It’s worth noting that the US train network is optimized for freight and we carry much more freight than many countries by rail.

High population density corridors like BOS-NYC-PHL-DC and SB-LA-OC-SD make a lot of sense and would be great to see even better service. We are seeing a high-speed, private service building track between Ontario, CA and Las Vegas. Distances to Vegas and Phoenix are short enough and population densities high enough that it makes travel by train from SoCal competitive with driving and air.

Back in the 2000’s we built a complete network operating system on top of QNX for OTN based long haul communication systems. In those days we had to sign SLAs on equipment and customers fined us for downtime. QNX was bulletproof despite running on our then PowerPC based custom CPU complex.

This is cool and it amazes me how the “big” home security companies don’t have this. My Ring cameras false detect all the time. The front will detect the flag as a person when it blows in the wind. The rear gets triggered by the pool robot skimmer. As much as I’d love to try this, I don’t have the time. It would be great though if it was built into something I could buy.

If you think of it, what is Meta going to do with AI? Their business model is ads. Maybe AI can improve ads some. For Meta, how does the billions spent on AI impact their bottom line? The reality is that if AI training requirements keep scaling based on Chinchilla scaling, there is going to be massive consolidation in training due to the scale that’s going to be required. Not everyone is going to be comfortable with Zuck running nuclear reactors. Already AI models are the fastest depreciating assets ever created. Moreover Meta is behind the AI curve so they are desperate to catch up. But desperate for what?

The supply chain is already dispersed, even outside Taiwan. Particularly as we move from single die devices to MCM, many processes are outside of Taiwan. JCET is in Singapore and Amkor is in Arizona and Korea for example. There is some cost to the logistics but it’s kind of in the noise on a per device basis. The cost is in the processes themselves. It is a gigantic pain to manage but it doesn’t not add such a high variable cost.

Semiconductor companies need gross margins of around 65% to grow and be able to invest in development of the next node. So this large additional variable cost really can’t be shrugged off as you suggest. If so, Ms. Su wouldn’t have mentioned it at all.

The overhead cost of a fab is fixed. So hard to understand why that would have such a wide variance. It may be true that the facility hasn’t been fully amortized so in principle it’s more expensive to make chips there. I can understand it being more expensive for many reasons. However I wouldn’t expect the cost difference to have a large variance. 5-20% is a very large range if the yields are comparable.

The article doesn’t say why the chips have a cost difference. The wafer cost of advanced nodes is ~$30k per wafer. Is the wafer cost different or is the yield different and hence the reason for the variance of 5-20%? All else being equal (same die size/design on same process) I suspect that a large part of the cost difference is yielded cost due to maturity of operations at the Arizona fab. Taiwan has had many years to optimize operations. You see this for any product initially when it moves to a new production site.

You need look no further than the poster child of red tape delayed construction projects: California High Speed Rail.

https://www.nytimes.com/2025/05/04/us/high-speed-rail-califo...

If you think how much of our collective lives the shoebomber wasted making everyone take their shoes off, I’d say that was terrorism enough.

For any fire, once the fire gets hot enough, it can be difficult or impossible to extinguish.

The fundamental problem is that battery fires get to be very high temperatures 1200C and cannot be extinguished at that point. I think the distinction you’re making about presence of thermal runaway or not is really rather irrelevant because yes you can put that fire out. That’s not the problem. The problem is that the devices do runaway and when they do it’s very difficult to put them out.

The ship in the original article was abandoned because it the fire could not be extinguished. The battery fire at Moss Landing could not be extinguished for 2 weeks.

Here’s a great video of the MountainView Fire Department talking about the difficulties of putting out EV fires. They explain that they’ve had cars catch on fire again 6 days later. They purchased new specialized equipment but at the time their department was one of the only fire companies that had this in California.

https://youtu.be/lnLqJqAT48E?si=UmEmSu841WyGU8PP

I have no idea what this says. Quite literally every single English language reference returned by a simple search refutes the assertion that EV fires are not more difficult or toxic than a regular car.

The fact that EV fires are more difficult to extinguish and produce substantially more toxic byproducts than regular gasoline powered cars is well established.

The toxicity of battery fires is widely reported: https://www.nytimes.com/2025/02/10/us/california-battery-pla...

Last month, a battery-storage plant went up in flames and burned for days, prompting the evacuation of more than 1,000 residents and shutting down local schools. The plant, located in Moss Landing, an unincorporated community in Monterey County, is the largest facility in the world that uses lithium-ion batteries to store energy. Residents have reported feeling ill, and many of them worry that the fire polluted the air, soil and water with toxins.

“Now you don’t see anybody walking outside because it’s terrifying, everything that’s going on,” said Esmeralda Ortiz, who had to evacuate from her home in Moss Landing after the plant began burning on Jan. 16.

This is fundamentally incorrect and even a basic Google search returns:

EV and gas-powered vehicle fires pose different challenges to firefighters. EV fires are harder to extinguish due to deep-seated heat in batteries, potential for reignition, and release of flammable and toxic gases. Traditional firefighting techniques might not be sufficient, and specialized equipment like submersion tanks or specific types of foam are often needed. Here's a more detailed breakdown: EV Fires: Higher Temperatures: EV fires can burn hotter than gas-powered vehicle fires, reaching temperatures of 1,200°F or more. Reignition: EV batteries can reignite hours or even days after a fire, even after the flames have been extinguished. Toxic Gases: Burning EV batteries release flammable gases (like hydrogen and methane) and other hazardous chemicals, posing a threat to firefighters and the environment. Difficult to Extinguish: Traditional firefighting methods, like using water or foam, may not be effective in cooling the battery cells to the point where the fire is extinguished. Special techniques like submersion tanks or specialized foam are often required. Prolonged Fire Suppression: EV fires can take longer to fully suppress compared to traditional car fires. Gas-Powered Vehicle Fires:

You made an account just to respond?

Generally it’s a good idea to sell what’s good about your product and not say things that are not true about other peoples products.

Also I wouldn’t brag about Kaiam.

It’s not correct to say that lasers are unreliable. Last year more than 20M transceivers shipped. Your statement is not at all supported by real field failure data.

The reliability of micro LEDs. and specifically GaN based micro LEDs is however an open question.

In the absence of any dislocation failure mechanisms it will depend on the current density and thermal dissipation. And just like any other material, it will have to survive in a non-hermetic environment and in the presence of corrosive gasses (an issue in data centers).

To get the 10G, it’s probably kind of like a VCSEL without the grating and so current density is probably high. How well you’re able to heat sink it is going to determine how reliable it will be.

Overall I like the idea. It looks like the beachfront could work. I’d spend more time talking about and how the electrical connection works and what kind of interface to a chip would be needed.

I’d also be careful before throwing shade on laser reliability because it could backfire on you (for all reasons above).

Quantum computing is still a technology of the future. When we are still talking about 12 qubits as a breakthrough, there’s a long way to go. Optical interconnects are the least of quantum computing’s problems.

However, it’s not correct to say lasers are unreliable. It’s fundamentally false and it’s not supported by field data from today’s pluggable modules. 10’s of millions of lasers are deployed in data centers today in pluggable modules.

It’s also useful to remember that an LED is essentially the gain region of a laser without the reflectors. When lasers fail in the field, they fail for the same reasons an LED will fail; moisture or contamination penetration of the semiconductor material.

An LED is not useful for quantum computing. To create a Bell pair (2qubits) you need a coherent light source to create correlated photons. The photons produced by an incoherent light source like an LED are fundamentally uncorrelated.

I would love to try this. I couldn’t tolerate CPAP. I used a dental appliance and that worked really well for a while, although was not perfect. I’m now on a path to have jaw surgery to move my lower and upper jaw by 1cm. I’ve started with extensive orthodontic treatment and it’s super uncomfortable. The thought of the surgery is also fairly terrifying. So if there’s a pill that could help, I’d sign up in about a millisecond.

If you’re Microsoft you may just want to give customers a choice. You may also want to have a 2nd source and drive performance, cost, etc… just like any other product.

In the 2 mixed reviews the hosts substantially misrepresented the property and the legality of their listing impacting the quality of our stay and the value we were receiving.

As someone that has a backbone and isn’t afraid of the consequences of reporting when I get ripped off, I find mixed reviews to be valuable.

Maybe they do now. I have only left 2 mixed reviews. One when the place they rented was not the one showed in photos. The people weren’t allowed to rent in their building and this was not something that was advertised. This meant none of the building’s amenities were available to us despite being advertised. I ended up getting dinged somehow. I thought maybe they could see the review by a 2nd account or something. Could be wrong.

Basic problem with Airbnb is that you’re dealing with small proprietors and the quality and value can be pretty inconsistent. Airbnb has relied on reviews to enforce this but what I’ve found is that “hosts” usually wait to file a review of me as a guest until I file my review. So there’s a disincentive to leave a mixed review.

Additionally there’s a creep factor in the number of cameras on the property. Hotels have lots of cameras but you don’t get the same sense that you’re being policed. I realize some of this is necessary but it can still be off-putting; usually everyone in the rental comments on the cameras.

Airbnb could normalize the value by enforcing standards and capping certain unreasonable charges in particular cleaning fees. A uniform cancellation policy would also help.

Additionally there are no rewards for booking Airbnb and no perks at all for repeat customers.

I’ve moved from Airbnb to Marriot and I get 4pm late check out, upgrades to suites, free breakfast, priority booking etc… and I don’t have to take the garbage out, bundle up sheets, do the dishes, etc…

If you’re making very low end HW maybe this is true. Because HW is something that you put into the real world there are other constraints such as power, cooling, space, security of supply, ability to ramp, cost, reliability, etc. The calculus for HW selection is much more involved than simply SW. Good SW/FW can be performant on much less capable HW but it does mean that SW engineers need to understand more about the HW. This is a very rare skill in 2025. Most SW engineers I’ve encountered cannot explain stack vs heap. Furthermore even fewer understand how to use malloc correctly.

The main problem with renewables is their capacity factor (amount of time they can produce their max capacity). In the US this is ~24%, in Germany I think it's ~12% (can be wrong here). The reason for nuclear here is that it has the highest capacity factor of any form of energy (see EIA.gov).

I think you missed the forest for the trees. I did incorrectly cite GPT-4 as I was going from memory and that's suspect sometimes. I also didn't elaborate and maybe I should have given the snarky comments I'm seeing.

Actually the amount of power available matters because you are consuming energy in time. If I have a 1MW plant and a battery, I can generate 1GWh in about 3 weeks. This seems a little silly though. A Hyperscale DC campus is ~150MW to 200MW. If you plot the larger ones, they are almost all near power stations with >1GW capacity (not all).

The industry trend is towards building 1GW datacenters. Last I checked these would consume ~8.7TWh (assuming PUE of 1). However, the 8.7TWh while relevant is meaningless unless the power to the DC can be 1GW. Since the plant itself has to generate more than 1GW (the plant has a cap ratio so more than this, plus other demand, etc..) for such a site, then it follows that there are limited number of sites in the US (this is public info see EIA.gov or Wikipedia).

Grok3 is already at 140MW (100 days of training ==> 336GWh) at ~10^26 FLOP. Model FLOP is increasing at ~5x per year so by 2030, we are expecting to be ~10^28 and that would take ~10GW (24PWh). If I am optimistic and say that the efficiency can improve by 1.3x per year, then we still need a very large generating station to meet the demand or we need to distribute among many smaller sites.

You can push the numbers around however you like but the conclusion is the same, the timing may be different.

There's a reason why all the hyperscalers are investing in nuclear, large generating capacity and the highest cap factor of any form of energy.

My 2nd comment still stands, and you left unaddressed (remember the forest?)..

It took about 1GW to train Chat-GPT4. If you look at the locations in the United States (>70% of all AI is in the US), there are only ~63 geographic regions you could put a 1GW data center. As AI models are growing at ~5x per year, it seems like the infrastructure is no in place to keep the AI models growing at that rate.

As companies like Google, Meta, and others look to nuclear power (it has the highest up time of any power source), I'm wondering how localities are going to react. Are people who are local to nuclear plants just going to be OK with these gigantic corporations consuming all this power in their backyard with no benefit to them while they take all the risk and impact of that power generation? I'm also wondering how these companies are going to deal with the excess nuclear waste. Ultimately it won't be Google or Meta dealing with the waste. How do we ensure that all the nuclear waste from AI is dealt with responsibly?

For the tire you can also use a penny. If you stick the penny in the tread with Liconln’s head down and his hair isn’t covered, then you need new tires. No AI. ;)

Partially correct. A private group is building high speed rail to connect Vegas to the Inland Empire. Construction has started and expected to be complete in 2028.

https://www.dot.nv.gov/projects-programs/transportation-proj...

https://www.brightlinewest/

Sure for InGaAs. The paper is claiming faster for AI so the relevant metric is to compare to what can be shipped today for such applications. To scale for AI we need devices that integrate with standard CMOS processes. More exotic materials can be integrated but they are not very practical for highly integrated systems. NVIDIA’s recent announcement for their copackaged optical switch is in a silicon only ecosystem.