joshuaellinger

Total Pipedream

Brian has been pushing a bunch of different projects in Central Texas since the Boring Company moved here. None of them have actually gotten any real consideration.

The only real tunnel that is likely to get built here is a transit tunnel in downtown Austin under the Colorado river.

My favorite is "Now, that's not a stupid idea".

I used to do that with a friend at my first job. We'd argue for hours until we found a good solution. The joke, of course, is that everything before then was stupid.

Of course, this assumes goodwill and friendship between the arguing parties. There is nothing wrong with losing an argument when a better solution exists.

I'm surprised nobody mentioned good old H. Ross Perot, the Texas billionaire who made Bill Clinton president in 1992. (https://en.wikipedia.org/wiki/Ross_Perot)

Perot peeled off enough libertarian voters from the Republicans to enable a Democratic victory. It odd looking back at him. He was not a proto-Trump exactly but he ran against the NAFTA free trade agreement in a way that looks a lot like how Trump ran against shipping all heavy industry to China.

My first reaction was that Yang is going to splinter the Democratic coalition thus enable Trump to get a second term. But maybe he can pull moderates away from the Republican side (especially in Arizona).

After looking at it for a second time, it looks really out-of-touch with reality and will not have a significant impact. Political polarization isn't a product of a broken political system. People are angry for a damn good reason. It's a reflection of all of the economic gains of the last 40 years going to the billionaires.

Someone has got to take control of the country back from the 0.1% like FDR did. That someone will likely be an unknown politician in the Democratic party who makes them pay for the costs of addressing climate crisis. Unfortunately, it doesn't look like that's going to happen any time soon.

I assume it was yours. Everything fits your description and I don't think anyone else is doing exactly the same thing.

I don't think the system is broken. It's just not an easy problem, especially if you can't use the market to set the price. California's only-when-you-sell system seems a lot worse.

Now, you could have a fairer system, but this is America and we're not particularly concerned about fairness when you get down to it. Austin's system actually has some progressive features that it is quiet about so the State does make them stop doing it. For example, if you are over 65, you can defer your property taxes until you die and accrue simple (not compound) interest. And houses have a quality 'grade'. Lower grade housing stock gets smaller annual increases which makes the system mildly progressive.

There is no state income tax so a lot of stuff gets funded through property taxes. It's not like the money is used to build bombs to blow up people on the other side of the world. Taxes are the price of a civil society and gaming the system was one of the primary causes of the decline of the Rome Empire.

So I don't agree "paying the minimum taxes you can get away with" is the same thing as "being screwed". It literally is not worth my time to protest most years, or even to hire a 3rd party, and I'm okay with that. I see companies like yours serving to keep the system relatively honest. The only thing I would change is that I'd add a $20 filing fee which is refunded if they make an adjustment to cut down on the junk protests, which are paid for my property tax dollars.

I live in Austin TX and got one of the mailers last year. I was wondering how you pulled it off.

The handwriting on the outside does make it look personal but, once you open it, it is obviously computer generated due to the consistency of the writing. If anything, I'd say the writing in inside undermined the message because nobody is going to handwrite a letter to solicit that business. You should test only doing the signature only and see if impacts your response rate.

We also got a couple of 'we want to buy your home' mailings. It does feel like it impacts our open rate but I can't tell if it is just due to novelty.

I think the main reason it 'works' as a business proposition is that the City is doing blanket increases and relying on protests to catch their mistakes. They do this (in part) because they are forbidden from using actual MLS data. It costs me 3-4 hours to do a protest so outsourcing it makes sense if your time is valuable. But when it came down to it, I wasn't willing to let them represent me -- something about it felt off -- and I wound up just accepting the increase.

The system doesn't feel broken -- just wasteful. How else are you going to do a property tax? I do wonder if this goes the way of medical billing (overcharge at first because you know you are going to lose some to protests) or it creates a watchdog effect that make the City limit the increases to an amount that is not likely to draw protests.

It might not work but.. I was suggesting that you try to do the reverse of what they are doing.

Large US companies look to other parts of the world to reduce their labor costs. If you try to justify the cost of the contract with them on your costs, then it would benefit you to use the higher US developer costs rather than the lower rest-of-the-world costs.

But, to have that work, you have to have come up with a reason that you need to be in the US.

It is probably easier to just sell them on hiring extra people in the place you are but you might get them to pay for you to move your business to the US.

Reading the comments, I noticed that my advice is very US-centric. The core still applies but you want to approach pricing differently.

I would set your pricing around a proposed move to the US so you can ask for money at US labor prices. If the move doesn't work out, you've got the contract priced right at least. Or I would inflate the number of heads you think you need to get to the same place.

Make it low-risk for you. Make it worth your time even if the contact disappeared overnight because your sponsor gets fired. The easiest way to do that is charge enough that you always have a year of expenses in the bank.

The key point is to justify your costs against their costs to replicate and then get a contract that doesn't pay you for your time.

I was in a similar position a decade ago. I took the money and it worked out for my ambitions (a nice living running a small company). It is up to you to decide which path to pursue.

In my case, there was an intermediary company and they loaned me the money needed to build the product and took it out of the future revenue. But a large company can advance you money or help you get a loan. Your job is to come up with a plan to do what they want and a price that makes it worth your time and ensures that you are around to support them in the future.

Some things come to mind: 1. It is hard to keep the product generic if you are serving a large client. I couldn't do it. It might be easier for you if the product lends itself to selling 'seats'.

   2. Find out how established your sponsor is at the company before investing too much effort in this deal.
   Have they made purchasing decisions like this?  Who would have to sign off on this deal?  Does it help 
   them with one of their top 3 goals for the year?  Just because you are talking to someone from a big
   company that doesn't mean they can actually write you a check.

   3. Get a copy of their MSA (master services agreement) and make sure they not expecting ownership
   for derived work.

   4. Everything costs x3 and takes x2 as long as you think it will.  Always does.  Plan accordingly.

   5. Don't hire a business guy to run the business.  It's your baby -- change the diapers yourself.  It 
   really isn't that hard and you can hire consultants to do finance and legal.  If you need sales, 
   think lead-gen.  You personally will close every sale at this scale.

   6. Hire someone who's entire job is to do client management.  Get someone good and pay them a lot.  
   Build that into your price.

...

Assuming you want to go the small business route, figure out what you think you need to run the business if they were your only client. Convert the roadmap into full-time heads and estimate what it would take for them to hire about a generic consulting shop to recreate what you are doing from scratch. Make sure each head is $20K/mo. Add 50% as a buffer. Your biggest risk is underpricing.

Then tell that you'd love to focus on them and it would cost you that amount to do it. Walk them through it. Talk about how it is less risky and faster to partner with you. See if it lines up with what they are willing to pay.

They may ask for ownership. If they do, ask why they want it? Being the largest client gives them more control than a 10% stake. If they insist, offer a 10% stake for the price you quoted to build it and reduce the ongoing price by 30% because the investment replaces the 'buffer'.

...

And a few more business/practical things that you might not know:

    1. You need a C corp if you want to raise money.  C corps are controlled by a board.  An LLC is controlled
    by a managing partner.  You want one or the other and you don't have to spend a bunch of money to set it up.

    On one business, I used generic paperwork from Techstars, Carta to manage the ownership, and a 
    registered agent to keep my government filings up-to-date.  Took maybe 2 days and costs < $1K/year.

    2. Once you get any scale, you have to pay yourself as an employee and report that as income.  I think you
    can get away with not claiming income at first but the IRS wants its share.  Use a payroll service -- do 
    not write a check for labor directly, even your own.

    3. If you have discrete chunks you can hand you, consultants can work.  Finding good consultants is hard.
    I personally favor direct hires because you get their full attention.

    4. Make sure to have all employees/contractors sign an agreement that gives you clear ownership of the IP.

Python has really extensive libraries. Here is the Nearest Neighbor one I've used in the past:

https://scikit-learn.org/stable/modules/generated/sklearn.ne....

...

Prior articles I've read point to R's startup time and memory footprint as disadvantages relative to Python. I don't have any experience with R so I'm not in a position to comment on it personally.

I would find an existing business that is hosting on the public cloud. Often, it is just a bunch of VMs that don't really leverage what the cloud has to offer -- they are just convenient.

Offer to migrate them from the cloud to a private hosting for what they are paying today but place them on hardware that has x2-10 the existing capacity. You provide all the IT work that they are paying Amazon to avoid.

I basically did that for myself. I was paying $10K/mo to Azure. To get the RAM I really needed would have pushed my costs to $20K/mo so I bought $30K (4 machines) in hardware and signed up for coloc @ $1K/mo but it cost me two months in my time to move everything. The time hurt more than the cash.

If you can do that for one company that is spending $100K/mo at Azure/AWS/Google, then your runway gets a lot longer.

Hard to take the analysis seriously when it (a) take UNIX as an example and (b) pretends Linux doesn't exist.

The Linux kernel looks to be a successful Bazaar project with a single person (Linus) having nominal responsibly.

I recently return to UNIX after a 20 year break. Linux in 2020 is much better than Solaris in 1998. However, Windows made giant leaps as well. Both still have massive amounts of deadwood but Windows does a better job of hiding it.

It's friendly, laid back but there is plenty of startup activity. (Power situation is better than SF in general because nothing is burning around here.)

Culturally, the vibe of SF is that everyone must to be exceptional at something. It values being (or pretending to be) AWESOME. I feel like the people of Austin are more interested in living a good life. However, there has been enough transplants now that everything is more mixed up.

The core difference in tech is that there is not the same level of money sloshing around. Less big crazy dreams. Less competition to standout. More bootstrapping. People who raise capital either do it on much worse terms or they go to the bay to get funded. Maybe rewind the clock by a decade or three (but not two) and it starts to look like SF.

Everything costs a lot less but you get paid less as well. I think you come out ahead if you have a family because you can actually afford to buy a house.

Also, it is pretty easy to work for an SF company from Austin if that's your thing.

Once COVID is over, easiest way to get a feel for the tech scene is to come to two to the bi-annual startup crawl organized by the Capital Factory.

Only take VC money if you want to create a large company and sell it in a 5-7 time frame. Or go out of business trying...

The main reason you should raise seed money is because (a) you need it to build the product and (b) you are worried someone else could take your market and need to move faster. Otherwise, build the business (as opposed to the product) and then raise 7-10M to expand it.

A common pattern is that you give up 20% of the company and a board seat then spend through the Seed round much quicker than planned with revenue lagging expectations. Then either fail to raise the A (happens most of the time) or have to give up another 20% and control of the company.

The control issue might be subtle too -- one board seat for the seed, one board seat for the A, two for the founding team BUT the money brings in a new CEO who has the 5th vote. The board actually controls the company (even it you own a majority of the stock) and the CEO will side with the money when it matters.

I would prove your economic model (will your current users actually pay for it?) before taking any money. In fact, probably the first thing to do is just ask them (see https://blog.asmartbear.com/vetting-startup-ideas.html).

...

My experience was that we raised $700K because we needed to cover the business guy's salary based on a trial by a single large enterprise client and couldn't get anyone else to adopt. The investors were great but it would have been better for all parties if we'd done something else with our time and their money.

My final bit of advice is put your revenue and growth expectations in a spreadsheet and game out the no-money vs with-money scenarios. Then increase your expense numbers by 20% and cut your revenue growth curve in 1/2.

I just implemented both a CSV parser and an address standardizer in numba (both CPU and GPU) running in parallel feed through a message queue with a bunch of workers subprocs.

It takes a bit of getting used to but the performance gains on impressive. Basically, my bottlenecks shift from compute to i/o.

I think you have to balance it against writing in C/C++. Mentally, it is basically the same work as writing in C (you manage memory/you write complicated for-loops) but you have good array support with numpy. The primary advantage for me that everything stays in the python runtime environment. You just run the code without any extra steps.

...

What is missing from the timing type 'toy' benchmarks is an understanding that there is typically more than one bottleneck in a real problem and it is easy to choose the wrong one to optimize and get little gains.

After starting C (30 years ago now), spending a long time in C#, then switching to Python a few years ago, I think the unappreciated advantage of python is that I have to abandon all pretense of caring about speed and just get stuff working. It basically solves the pre-mature optimization problem for me by being a fast interpreted language rather than a slow compiled language.

And... it is back (for us at least.)

This is great. I suspect that I have exactly the right use case for this.

The two main issues with running SQL Server are (1) you have to license all the cores on a system and (2) the standard license only recognizes up to 64GB RAM. So I actually wound up buying a 3GHz single-socket system for around $10K to save $20K on the SQL license.

With this, I can move a couple of the big DBs to another system that has 32 Cores with 256GB RAM and the entire DB will fit in memory, put in 5GB ethernet, and gain a tremendous amount of performance.

But, more importantly, I can migrate the workload on a case-by-case basis. Human costs always dwarf my software and hardware costs.

I read both sides and it looks like The Intercept really doesn't have a leg to stand on. Greenwald offered to publish the story elsewhere (which he is explicitly allowed to by his contract) and they said no. He published both the article and the emails leading up to his resignation. The Intercept clearly breached the terms of his employment contract and expected to pressure him into accepting it.

I think what makes it hurt so much for Jeremy (who strikes me as good guy) is that he is still loyal to the organization and doesn't want to face up to its hypocrisy. I don't know how he reconciles his opinion of Betsy with her comments accusing Glen of acting like a spoiling child.

I want Biden to win and think the NY Post stinks but Greenwald is right on the merits.

Sounds a lot like Alan Moore's complaints about comics.

But I think you assign problems to the game that are really problems of the audience. Who wants to be a random character in a Game of Thrones campaign? How exactly is it fun to get knocked out every 2-3 sessions and have start over?

Some people do play that way. It is really easy to make D&D 5e deadly. The rules are a lot less forgiving when you play them as written than you paint them. Sure, you get death saving throws. But once you are down you have to make them every time you get hit.

I run a campaign for a bunch of 10 year old kids. D&D is just the right balance for them.

I wrote something like this but used a GPU to do all the heavy lifting. It was mostly just to learn how Numba's GPU support worked so I never benchmarked it. It was very fast.

I think most people don't appreciate that CSV parsing is actually a compute-bound problem because the rows and column positions are known in advance.

Overall, I'm not sure the complexity justifies the performance gains unless parsing is in your critical path.

Military personnel are trained to be apolitical much more than you might realize from the outside. My information is from a cousin-in-law who just retired.

If a few generals who sided with the president, they would face a lot of resignations all the way down the command change.

In Germany in the 1940s, Hilter got control by literally assassinating the generals who opposed his takeover to take control. But even that would have likely failed without the massive economic crisis caused by the punitive settlement of WW I and without the unusual characteristics of German society. (no, Trump is not Hilter and the US in 2020 is not Germany in the 1930/40s.)

We have tested it a couple of things in the US. Truman recalled MacArthur for attempting to publicly influence political opinion. I forget about the others.

So... the mostly likely thing if it is close is that it takes a while but the political establishment throws it to Biden, Trump sulks into the sunset (maybe even prison eventually), and the Republicans don't care because they have just got a 6-3 on the Court for a generation.

Take a look at Numba for JIT.

The primary drawback of Numpy alone is that you pay the memory-to-cache latency for every operation. Numba compiles into C on the fly. It makes Python fast when it matters.

The dealership model comes from the abuses by the OEMs during the first part of the 1930s. They would sell you cars and then charge a ton for repairs and lock you in to only getting repairs from them.

The local dealerships are small businesses incorporated in the states so they can be subject to more political pressure and local control.

It has been eroded over the years but originally the dealerships protected consumers from the sort of bad behavior you see from Telsa and Apple.

Not that I love the dealership model but there is real competition at the local level on car repair in a way that there doesn't appear to be with Telsa.

Render consistency.

I was doing automated image capture of some data sites and diffing them to see if they changed. Chrome would jitter. Firefox drew the same bits every time.

Weird -- it appears to be missing the last 10-20 years of development in random numbers by the data science and cryptography community. They have put a lot of work into it.

If I recall correctly there are two major angles, 1. there is an instruction in modern Intel chips that samples random thermal noise. 2. there are a whole class of elliptical curve approaches that pass a bunch of randomness tests.

And I'm just scratching the surface here.

GPU servers and coloc are pretty cheap these days. $1K/m rent per 20A of power. ROI on hardware is usually 3-4 months max (ie - for the cost the machine at AWS for 3-4 months, you can buy the same thing).

Lead time might be a problem for you but you can probably do it in a under a month if you take available stock at your vendor. I work with a company called PogoLinux (http://pogolinux.com) out of Seattle and they sell boxes that have 4 GPUs in them.

That said -- the other advice is right. You can probably get by with a much simpler model. The coloc route would probably only be better if you are can't change the models due to people constraints and the ML stuff doesn't have a lot of AWS dependencies. SysAdmins are a lot easier to find and hire than ML specialists.

Why are you comparing defense spending to a pension fund (SS)? Hardly seems appropriate. Having a bunch of old people around is expensive. Are you suggesting that we should go back to letting them starve when they get too old to work or die young because they can't afford medical care?

Better comparisons are Police ($100B) and prisons ($80B) but really you should be looking at spending as a percent of GDP. Defense is about 3% of GDP.

So we spend (nominally) 7% of our GDP to take care of our elderly. We spend (nominally again) 3% to police the world. No other country comes close to us in military spending but we probably make more from our empire than it costs for us to maintain it. At least so long as we don't get into too many stupid wars (cough cough Iraq).

I can see what we get for SS and Medicare and I'm fine with the taxes to support it. I'd feel a lot better about the military spending if less went to crazy boondoggles like the F-22. It feels like a giant waste -- we shipped all our factories to China and none of this crazy fancy tech is going to help us if we get in a fight with them.

Palantir feels like it might be another boondoggle but I've seen a bit of it at a large company I work with and it actually solves some real problems. It appears to be a slick integration of Spark and Git. It has some potential to solve problems that they haven't been able to touch in the 10 years I've been contracting for them.

I'm in the midst of returning from the cloud to coloc.

It's looking like for about $50K in hardware I get x10 the ram/compute/storage/(internal) network I get from the cloud at around $10K/mo. However, it is taking me about $25K in labor to setup. Hosting costs about $1K for a rack and a decent pipe to the server.

So more than a hobby but not approaching real scale.

The main barrier to using a lambda model is if you have anything that smacks of a DB. It would take me $100K+ to transition away from SQL. If you can do lambda w/o a big retool cost, then it is probably viable. If you are just running a bunch of VMs on the cloud, it is pretty expensive.

...

Interestingly, my main client (a very large company) just went the other direction and moved all its compute to a system that is Spark underneath running on Azure. They are trying to decommission some expensive TerraData instances. So far, it is a mixed bag -- it is a big step forward (for them) on anything that is 'batch-oriented' but 'interactive' performance is dismal.

Price appears to be a big motivation. I always forget that most large enterprises run on exotic stuff with crazy service contracts that makes the cloud look cheap.

1. Nothing I've found comes anywhere close to Emacs support for keyboard macros. Zero latency to record. Records every keystroke. Very very consistent behavior. Infinite undo. Kill Ring. Register support.

Despite being an expert Emacs user, I'm mostly in VS Code these days because it has syntax highlighting, git support, and linting out of the the box. But I always go back to Emacs when I need to transform a bit of text.

2. Comparing two text snippets, as opposed to files.

There are a small set of use cases where you want to look at every position where something has changed. You can open two buffers in Emacs and use compare-windows to walk through them. Since the IDEs are all file-oriented, they don't work for this case well.

On the smaller end of the scale, I have a $12K/mo spend with Azure. I decided to go back to Coloc.

For under $50K, I have 4 machines with an aggregate 1TB RAM, 48 cores, 1 pricy GPU, 16TB of fast SSD, 40 TB HHD, and infiniband @ 56GB/sec. Rent on the cabinet is less than $1K/mo. It's going to cost me about $20K in labor to migrate.

So the nominal break-even point is six months but the real kicker is that this is effectively x10-30 the raw power of what I was getting on the cloud. I can offer a quantitatively different set of calculations.

It also simplifies a bunch of stuff: 1. No APIs to read blob data -- just good old files on an ZFS share. 2. No need to optimize memory consumption. 3. No need for docker/k8s/etc to spin up under load. Just have a cluster sitting there.

There are downsides but Coloc beats the cloud for certain problems.

Any plans for an Azure Blob API?