#3 The Startup Party: reflections on the last 20 years, what could replace the Tories, and why

What did I really spend my time on in No10, what does the VL plan imply for the old parties 2024-30; more on AI/No10

Pessimism of the intellect, optimism of the will

It is not necessary to hope in order to persevere

UPDATE 10/6, two years ago exactly I started this substack. Yesterday afternoon I posted this blog and suggested, parenthetically, a targeted campaign to push Boris out of his seat. A few hours later he issued his laughable resignation statement.

Why has he gone?

In 2020 I watched him crack up. He broke his deal with us. In our last conversations he was deranged, ranting about getting the Cabinet Secretary to give Carrie a job ‘with Kate to get her out of our hair’ then ranting that ‘I’m the fucking Fuhrer’ and he was sick of me telling him No — whether it was his mad bridge to Ireland, secret and illegal donations for his gold wallpaper, or his ‘money’s just digits in a computer’ attitude to taxes and spending. In spring 2021 he started lying about covid claiming there’d never been a Plan A for ‘herd immunity’ by September and therefore no shift to Plan B. Given herd immunity by September was the official DHSC plan, briefed as such by the government and personally by Hancock and senior officials across the media and explained in thousands of pages of government documents, this was, even by his standards, extraordinarily bad and stupid.

I and others always knew he was not fit to be PM but in 2019 we thought trying to control him was the best of a bad job (cf. below). In 2021 we decided to remove him and started a concerted effort to put him under constant pressure so he’d blow himself up. We were helped by officials in No10 and 70 Whitehall appalled by his behaviour. We were mainly helped by his combination of pathological lying and surrounding himself with courtier-fools.

He clearly had no intention of trying to honour the promises we’d made in the 2019 election. He dismantled or let Whitehall stop a large amount of what we’d started building in 2020. Practically everything in the media about what we were trying to build in 2020 is rubbish, if you’re interested in what we actually focused on and cared about then look below. Later in the year he broke a core 2019 promise not to raise taxes. Tory MPs disgracefully supported him. He now burbles, like many Tory MPs, about ‘growth’ and ‘lower taxes’. He was the one who stopped the plans for growth and lower taxes.

He’s trying to pretend he has been pushed out ‘as revenge for Brexit’. No. We pushed him out because he betrayed the Vote Leave plan to do Brexit seriously and started lying about why over 100,000 people died, most of them unnecessarily.

He’s trying to pretend Sue Gray was a conspiracy. No. Appointing her was a perfect sign of his uselessness in No10 and his ignorance of how Whitehall works. I wrote about how Tory MPs didn’t understand Sue Gray in 2014. His pathetic claim that appointing her was a sign of his ‘faith in the impartiality of our systems’ is a perfect example of how over 25 years Tory MPs have become totally unserious about real power and how Whitehall works.

He leaves disgraced and ranting nonsense.

The Tory Party, also disgraced, should be closed and replaced.

---

#1 here.

#2 here.

The third looks at:

• Why did you go to No10 July 2019?

• Why did you want a ‘rubbish Cabinet of Yes Men’?

• What really happened with Sajid Javid resigning and Sunak?

• The media story was you spent your time on politics, the media, ‘campaigning’, ‘culture wars’ in 2020, how did you actually spend your time?

• Why was ‘a British ARPA’ so important?

• What was the truth about the shakeup of the government communication machine? (The Undertaker Plan)

• Truth about ‘hard rain’ (didn’t happen)

• What was your role in the election?

• How hard is it to deregulate? (Very very very)

---

AI again

In the last blog I discussed Marc Andreessen’s piece on AI and regulation.

Some of those who are more pro-regulation have had a go at it.

A few things they’re underrating:

1/ Brussels is a joke on AI regulation.

So far Brussels has taken the lead on attempting regulation on AI. Their proposals are laughable. So laughable that they will actually force the main US/UK companies to withdraw many services to the EU. If we hadn’t done Brexit then DeepMind, Faculty and others in London would be scuppered. The EU is, post-Brexit, mostly irrelevant to the leading edge of the field because it lost London. The top companies are in California, China or London; none are in the EU. Note the news today that Palantir has just chosen London to be its European HQ. Regulation scenarios are a reason why. Having talent in London is crucial if you want the UK to help seriously with AI safety.

Brussels’ priority is a/ increasing the power of EU institutions, b/ attacking American tech companies. It puts out documents about the EU becoming the ‘most trusted’ jurisdiction on AI — not a technology leader but a leader on ‘trust’. (Leave aside the comedy of the Commission thinking it will be a global leader on ‘trust’ about anything.) It is similarly bad on aspects of genetic engineering. It was awful on vaccines — it was obvious when we looked at their plans in spring 2020 they were awful (hence why we did the Taskforce), and the now-German PM said accurately the EU system was ‘really shit’ and ‘a disgrace’ and Germany could not ‘let this shit repeat itself’. And the Charter of Fundamental Rights (NB. this is NOT the ECHR) guarantees this problem gets worse and harder to fix, just as we pointed out during the referendum. A sad spectacle for the home of the Renaissance, modern science and the Industrial Revolution.

Across the Valley and tech world people are twigging what we said in the referendum.

(Good to see PG making the point but it wouldn’t be a ‘surprise’, we said exactly this in 2015-16.)

Brussels proposals are SO bad many in the Valley don’t want them watered down and want to see the EU cut off and other companies therefore fold so finally the EU gets a hard lesson and gets serious about technology.

Its failure over GDPR has been hidden, partly because the politicians and media don’t understand the issues. It told everyone GDPR was intended to hit Google, Facebook etc and persisted after it was pointed out that, in a classic example of unintended consequences, GDPR wouldn’t solve the problems it claimed to be aimed at and would help the people intended to be harmed. They didn’t listen.

Those who want wise regulation of AI should consider how bad EU action DISCREDITS the idea wise regulation has any reasonable chance given our actual political institutions and leaders. This undermines support for any regulation among crucial audiences because they update towards ‘any regulation will be a shitshow’.

2/ Our governments are a joke on known existing deadly serious bio-risks, so why think they’ll regulate wisely on AI?

There is a strong case that covid was a lab leak (natural or synthetic). I pointed out this danger many times over years before 2020. Have you noticed a great rush to improve biosecurity, which should happen even if covid was natural? No. Just recently in Sudan a lab full of god knows what fell into the hands of some murderous gang. Covid may well have come from gain of function research. And omicron may have come from it too! Have we stopped this work? No. Have we listened to George Church and Kevin Esvelt on all the risks they’ve identified? No. Has the US government lied about it? Repeatedly. And Fauci has been disgraceful on this subject and greatly undermined confidence in government honesty about research.

Look at our own government on covid. In 2021 it decided to shut down the actually world-leading sewage monitoring we built in 2020! It has taken or failed to take many actions 2021-23 that prove it is not interested in taking bio-risks seriously. I won’t go into all this again here but the basic point is simple: across the west governments not only performed terribly on covid but have strengthened the same appalling bureaucracies that failed so badly, and opposition parties and old media everywhere have been fine with this.

So those arguing for AI regulation are saying that the exact same authorities across the western world who did such an appalling job on covid, the people who have shown no interest in reforming the shitshows of CDC and FDA etc, who even after covid have just ignored the whole issue of gain of function and lab security, should be trusted to solve problems 100X or more harder on AI regulation which is moving rapidly up an exponential?!

I stress again, I’ve been pointing out dangers in AI for many years long before the subject went mainstream. I am NOT saying ‘ignore it, it will all be fine’. But those who think there are serious dangers — which includes Andreessen — cannot reasonably go down the path of ‘well we just have to hope this time is different and the same people and broken institutions will somehow pull a miracle out of their asses on AI’.

No no no.

3/ If current leaders and institutions panic into action they will do similar things to Brussels.

Left to itself Whitehall will do things that simultaneously add friction to people working on, say, beating cancer with AI but add no friction to the real bad actors.

We won the referendum in 2016. Seven years later Westminster has failed to … scrap the dumb cookie popups. Look at the Online Harms Bill. Panic hyped by the newspapers led to years of SW1 babble, No10 thrashing around, and there’s a joke Bill and SW1 totally incapable of discussing the issues sensibly. It’s so embarrassing and SW1 so incapable of discussing technology seriously there seems to be an implicit deal to stop discussing it. You think they’re going to regulate AI wisely?!

It’s more likely they will drive companies out of London to California than the old system works well on AI.

4/ CHINA.

If you really are serious about AI regulation you have to confront a fundamental problem that in America the national security establishment and much of Silicon Valley sees the real danger as China — not DeepMind, OpenAI and Anthropic.

Much of the pressure for government action has strengthened those who want to race faster. I.e those wanting to stop the race are in many ways accelerating it. Welcome to politics!

Many of those who are technical experts in this field are very naive about politics and this problem gets bigger the closer someone is to the EA movement. Do you think you will persuade the White House under Biden or Trump that we should kneecap top western firms then watch Chinese firms take the lead?

What’s your plan for cooperation with China given a) international cooperation can only work with Chinese agreement and b) powerful forces across the political divide want to escalate conflict with China rather than cooperate?

What’s your plan to convince US voters in Pennsylvania that the US government should unilaterally STOP pushing the frontier in AI in case it accidentally destroys humanity, choose NOT to push AI research for biodefence, cyberdefence, curing cancer, spreading wealth and so on, and voters have to accept the risk that China will acquire strategic advantage so if AI doesn’t kill us all then we’re all under Chinese control, and ‘this is great expected value, here’s the EA paper with the calculation’?! (Yes, I know Eliezer and yes I know the arguments on ‘don’t rush to AGI just to beat China’, that’s not the point.)

Bear in mind many of those trying to argue this are the people who for more than a decade have tried to suppress the global Terminator-as-AGI meme because ‘it’s stupid and misleading’. Do you think they will navigate this communication challenge?!

5/ Better safety will come from building not regulation.

There is a very strong argument, made by Andreessen in his piece and actually being implemented by Faculty AI in London, that real progress with safety is more likely to come from iteratively solving safety problems with more and more advanced narrow AI systems than from the governments we have trying to regulate.

If there turns out to be a solution to the biggest alignment problem — how to secure human survival if we build systems that a) are much more intelligent than us across all important domains and b) are connected to the physical world so can do manufacturing, engineering etc — then it will probably come from such iterative work. The idea that we a) try to freeze progress where we are now (extremely unlikely and anyway ignores that even without any further new ideas engineering based on current ideas will drive phenomenal progress) while we b) try to figure out what could contain a hypothetical GPT-2035 version (maybe as hard as making many Nobel-like breakthroughs together, nobody knows) is rejected by many at the edge of the field.   

I’m not saying this argument is right. I’m saying many people at the edge of the field including those responsible for massive investments, much vaster than the UK government or Brussels are making, believe it. If you don’t agree then you need to persuade them.

What is to be done?

Here I’ll just mention the UK.

• The UK government does not have the technical talent it needs to help think about AI including regulation.

• It does not have structures in place to do it. Structures we started setting up in No10/Cabinet Office have been blown up by Boris, Truss and Sunak. The Cabinet Office is a disaster zone. The NSS/JIO etc don’t have the people needed. DSIT has been a step backwards for British science. Whitehall has been engaged in a ferocious multi-month turf war over AI. (For the old media setting up a new department for X is a sign of X being a priority but in the real world it has shown key players X is not an actual priority and the PM doesn’t understand Whitehall).

• The announced ‘ambition to be a world leader’ by buying 3,000 GPUs is a proper Doctor Evil ‘one MEELEEON dollars’ announcement. There are individual hedge funds with 3-4 times more than this. This has shown actual experts that either we do NOT have an ambition to be a ‘world leader’ or we’re run by people who don’t understand what they’re announcing (or both). An actual plan needs roughly 30X more GPUs.

So we need a serious Taskforce which means inter alia:

1. Freedom from Whitehall HR so it can get great people. A normal thing will not be able to hire the people needed. ‘People ideas, machines — in that order!’ The most fundamental challenge is a) getting top people into government thinking and avoiding business as usual, b) persuading some of the smartest people in the world, people like Terry Tao, to spend some fraction of their time on AI in return for the government doing them various favours (e.g funding maths/physics projects they want funding). There’s a lot of talk about a ‘Manhattan Project’ but that had an enormous structural advantage: Hitler pushed most of the smartest physicists in the world to America and highly incentivised them to work on the bomb. So if you are serious about AI / safety you must find a way to get politicians who do sod all on gain of function to mobilise the smartest people to work on this.

2. The TF controls the compute budget and has freedom to start buying stuff fast (partly so it can enable safety research). The current compute budget is double-counted to cheat the media, it’s not under the control of the TF, it’s anyway in the next SR period(!), officials are lying to No10 about what is being procured, and anyway they can’t do anything fast because it’s Whitehall. SNAFU.

3. The TF has the authority of the PM to act. If AI sits in a normal government department you will get normal results. If HMT is in charge of ‘business cases’ you’ll get normal timetables and results. This is incapable of dealing with models becoming more powerful at maybe 5-10X per 6 months.

4. The TF can engage seriously with the deep state which is in a terrible position on AI and has dropped the ball, partly also because of usual problems like the country actually being mainly run by HMT officials who refuse to allow long term budgets because they think everything should be micromanaged by them. 28 year old HMT PPE-officials routinely squash funding plans from Nobel winners on the grounds they think ‘this isn’t a good bet’ — black comedy.

5. A serious Five Eyes plan as the first step towards broader international efforts including backchannels to Chinese intelligence. China wants more power. But China doesn’t want AI enabling dissidents to overthrow Party rule. Now there’s no discussion with China on many global risks because of the insane desire of so many to escalate war over a Chinese island full of Chinese people which WE have said for 50 years should be part of One China.

6. Either a) support a permanent manned lunar base to drive international cooperation and create new institutions for it, or b) come up with something else to do this.

7. Those who want regulation have to draft actual laws and an explanation for how they will really work and how they will cope with, for example, leading edge models rapidly shifting to open source systems that can be shifted anywhere on earth.

The DSIT and Cabinet Office documents floating around for months are the opposite of a serious plan and are another national embarrassment.

In a nutshell:

1. There are real dangers in AI but they are barely understood in politics world.

2. It’s very easy to create media panic then get politicians to pass regulation. Instead of doing this, those worried about AI should first consider prior, much harder questions. How to improve 10X the average quality of people in power who will actually figure out and decide these issues? How to create institutions that can handle AI 10X better than bio-security and nuclear security (cf. repeated crazy failures on nuclear bombs)? The same government people plus ‘let’s set up a CDC or IAEA for AI’ guarantees not just failure but bad unintended consequences. NB. the Manhattan Project, the ICBM project, Apollo, and personal computing and the internet were all world changing projects that relied on highly unusual and super-productive institutions. This is really really hard because all normal government fights to stop it.

3. Those worried about superintelligence have no plan for how to deal with China given politics-world wants to fight China more, not cooperate. Much of their activity is encouraging the US deep state to race faster with China which is the opposite of what they want.

4. The EU is a joke on the subject but this is impossible for many Insiders to absorb for obvious psychological reasons. They are hardly going to say ‘oh dear turns out Vote Leave was right’.

5. There is no Royal Road to aligning superintelligence. It is super duper unlikely to come from normal government. There’s no reason to think something this hard should be any easier than fundamentally changing how government works and plenty of reasons to think that a) solving the technical problem of aligning superintelligence and b) fundamentally changing how government works naturally connect as problems. Intelligence is power. Power is politics. Superintelligence is inherently revolutionary for politics. So stop looking for shortcuts comrades! If you’re going to solve alignment, there’s no escape from grappling with the issues in von Neumann’s Can we survive technology? The world is ‘undersized and under-organised’. Get building the lunar base!

What do I expect?

I expect the normal. Normal government, normal delusions, normal failure, Britain blows a huge opportunity that was part of the reason we did Brexit even though there’s a massive open goal because this is SW1-normal-mode. I expect the PM to sort of try to do this properly but not do it properly because doing it properly means actually controlling the government and all signs are he doesn’t want to face the blowback inherent in actually controlling the government.

Finally, imagine a No10 with staff that have long cared about the pros and cons of AI and saw it as a core priority — which actually existed in 2020 (see below)! The whole of Westminster will attack it as ‘mad’, ‘obsessed’ and ‘dangerous’. MPs, hacks and officials will mobilise to crush those trying to make progress. The officials who want progress will keep quiet guessing an attempt to take government seriously will fail as usual.

Remember when we started setting up a No10 data science/AI team in January 2020? Remember how Whitehall and the old media laughed? Remember how weeks later the government collapsed unable to handle covid data with me scribbling x2 x2 x2 on a whiteboard in the Cabinet room in response to faxes? Remember how most of what we started building in 2020 to fix this has been closed and the MPs and hacks have been happy with this?

Good luck comrades…!

Ps. Thinking about a leader for The Startup Party.

What are some characteristics to look for in a political startup aiming at fast growth and takeover in 2028?

• Globally famous, can make news with a tweet.

• Has built a globally successful brand.

• Track record of calm moral courage in the face of physical threats.

• Track record of taking the side of the median swing voter against the hysteria of London’s graduate and media classes.

• Track record of effective consistent message over years actually changing the dynamics of a big political issue — something that the Tory Party has failed to do once since 2010 and has no idea how to do or the attention span needed.

• Innovative use of technology for communication.

• Looks like could build and lead a great team.

I present — JK Rowling!

Now add people like:

• Marc Andreessen, Secretary of State for Technology and Build Build Build.

• Alex Cooper, former commanding officer SAS and hero of covid testing, SoS War and Defence.

• David Deutsch (Leave) and Tim Gowers (Remain), co-chairmen of UK JASON.

• Substitute some officials in the PM’s private office 2020 to ministerial jobs: Hannah, Imran, Alexandra, Jonno... In each case a dramatic improvement over the ministers either Tories or Labour could suggest and crucial experience in how power really works.

• Victoria Woodcock, Cabinet Secretary

You get the picture…

Suddenly you’ve made a transition from a political class exemplifying the worst of our country to a team that could take on anybody!

No, I haven’t discussed this with any of those named. This is a thought experiment!

Pps. Someone should run a targeted campaign to remove Boris as an MP wherever he pops up. I assume he will ditch his current seat and try to blackmail Sunak into a safe one. E.g set up a system so that students from all over the country can be matched with volunteer addresses so they can register to vote in his constituency and swamp the electorate. There’s lots that could be done. Don’t leave it to dud Starmer…

(This website has a collection of some brilliant pieces, from a profile of Frank Sinatra to Bret Victor! h/t Tyler Cowen.)