Hauke Hillebrandt (@Ha_uke) header
Hauke Hillebrandt (@Ha_uke) profile picture

Erstwhile researcher at @Harvard @UCL @CGDev @CentreforEA and Founder of Let's Fund.

London
Joined September 2018
Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)Photo by Hauke Hillebrandt (@Ha_uke)
Hauke Hillebrandt retweeted
Government imposed policy costs make up 57% of the clearing price of a new flat in London and 42% of a new home in the Home Counties. Despite a big housing shortage and positive planning reforms, new home starts in England are declining from 182,000 in 2022 to under 125,000 last year. While the macroeconomic context has worsened with build cost inflation and high interest rates, the Government is imposing more burdens on new homes, not fewer. Taxes, levies, planning, and mandates add over £440,000 to the cost of delivering a home in London and nearly £225,000 nearby, with several of these imposed recently like the Second Staircase, Future Homes Standard, Building Safety Levy, Biodiversity Net Gain and Residential Property Developer Tax. Individually these may be justified, but together they are making new homes unviable. As a result, large developers are buying less land and small builders are going out of business; nearly 4,000 construction firms went bust just last year. If we want to build more homes, we should be easing the burdens on them, not adding more. Making housebuilding viable again is the most important issue in housing policy and my co-author, Edward Donovan, has completed excellent modelling showing how big the challenge is. Read the full report: cps.org.uk/wp-content/upload… And simulate what changing the policies does with Edward's modelling here: easing-the-burdens.cps.org.u…
75
273
97
1,105
242,647
Does Anthropic's best internal model + $50M in data gathering budget + 3 months currently dominate any one particular human expert at any one particular task?
37%Definitely yes
26%Probably yes
18%Probably not
20%Definitely not
358 votes • Final results
12
3
3
19
5,701
Hauke Hillebrandt retweeted
A good amount of recent AI agent chaos comes down to reward hacking. People want to blame it on bad training environments, and that's often true. But we find >100 examples with no obvious env screwup. We manually tag all 610 as being attributable to such a screwup, or not.
9
14
2
127
5,592
Hauke Hillebrandt retweeted
‼️ Sehr eindringlich warnende Worte vom BND-Chef: Moskau stelle uns auf die Probe und wette auf unsere Schwäche. Denn Putin sei über den Punkt hinaus, wo er seinen Krieg einfach beenden könne, ohne sich selbst zu gefährden, sagt Martin Jäger bei der Geheimdienst-Anhörung im Bundestag. Sein Eingangsstatement im Detail: — „Deutschland ist in Gefahr, in einen gewaltsamen Konflikt mit Russland zu geraten. Der Anschlagsversuch in Leipzig ist nur ein Beispiel dafür. Von keinem anderen Land geht derzeit eine so direkte, umfassende und konkrete Bedrohung gegen Deutschland aus.“ — „Nicht wir haben diese Konfrontation gesucht: Russland drängt sie uns auf, sieht in uns einen Hauptgegner. Weil wir Russlands revisionistisch-expansive Politik und den Angriffskrieg gegen die Ukraine nicht hinnehmen, weil wir der russischen Einschüchterung unserer Bündnispartner im Baltikum und anderswo in Europa entgegentreten. Deutschland ist zu einem Haupthindernis russischer hegemonialer Machtprojektion in Europa und damit zur Zielfläche russischen Handelns geworden. Wer sich nicht fügt, wird zum Gegner. So lautet die Moskauer Lesart.“ — Dies bedeute: „Russland stellt uns auf die Probe und beobachtet genau, wie wir mit dieser Situation umgehen. Jedes Zeichen von Wankelmut lädt aus Sicht der Führung im Kreml dazu ein, den Druck auf Deutschland weiter zu erhöhen.“ — „Präsident Putin ist über den Punkt hinaus, wo er den Druck auf die Ukraine einfach einstellen könnte, ohne seine Machtbasis im Inneren zu gefährden. Er muss einen Erfolg in der Ukraine vorweisen, der die bisher erbrachten immensen Opfer aus russischer Sicht rechtfertigt. Das setzt ihn unter Zugzwang.“ — „In Moskau herrscht aber keineswegs Kopflosigkeit. Die russische Führung weiß um die Begrenztheit ihrer militärischen und wirtschaftlichen Mittel. Sie kompensiert eigene Schwäche durch Repression im Inneren, durch die konsequente Ausrichtung der Ressourcen auf die Kriegsführung und — das ist entscheidend, — durch Entschlossenheit und Risikobereitschaft.“ — „Mehr als Material, Geld und Menschenleben zählen der russischen Führung Willenskraft und Opfermut. Das sind die Währungen, in denen Russland handelt. Sie werden auch uns in Rechnung gestellt. Moskau agiert in der Annahme, dass wir russischen Druck und russischer Einschüchterung letztlich nicht standhalten werden. Moskau wettet auf unserer Schwäche.“ — Worauf müssen wir uns einstellen? Jäger nennt vier Punkte: 1. „Russland legt es darauf an, die Ukraine im kommenden Winter in die Knie zu zwingen. Wir müssen mit russischen Attacken von einer Härte und Brutalität rechnen, wie wir sie bisher so in der Breite nicht gesehen haben. Der Krieg in der Ukraine tritt in eine neue, überaus gefährliche Phase ein.“ 2. „Russland versucht zweitens, westliche Produktion und Lieferung von Waffen und Wehrmaterial womöglich zu unterbinden, auch kinetisch, auch bei uns.“ 3. „Drittens: Russland zielt darauf, Europas politische Unterstützung für die Ukraine zu brechen. Wir müssen uns auf Einschüchterungsversuche und weitere Einmischung einstellen. Vermutlich wird es auch erneut nukleare Drohungen geben. Dem müssen wir standhalten.“ 4. „Aktuell sehen wir, viertens, keine Hinweise auf einen groß angelegten Angriff Russlands gegen Nato-Gebiet. Allerdings könnte es, insbesondere im Baltikum, zu niedrigschwelligen militärischen Aktivitäten Russlands kommen. Und in diesem Kontext lassen sich auch False Flag Operations nicht ausschließen.“ Zusammenfassend: „Russland führt einen Schattenkrieg gegen uns“, sagt Jäger. „Dieser unerklärte Krieg ist nicht mit herkömmlichen militärischen oder rechtlichen Kategorien zu fassen. Wir befinden uns nicht im Verteidigungsfall und werden trotzdem angegriffen. Ein Schattenkrieg eben.“ „Die russischen Geheimdienste agieren hochprofessionell und rücksichtslos. Die Gefahr von Anschlägen gegen Ziele in Deutschland ist weiter real und gegenwärtig. Dem müssen wir uns entgegenstellen.“ Der BND sehe sich als „erste Verteidigungslinie“ (Verweis auf Geheimdienstreform mit neuen Befugnissen).
129
764
96
2,031
311,166
Hauke Hillebrandt retweeted
Something I noticed about working with Fable...
173
503
142
7,454
365,157
Hauke Hillebrandt retweeted
Interesting that the version of Mythos 5 in this incident is trained on the Constitution but lies/gaslights the Github maintainer to get them to accept the malicious PR anyway. Points for the Yudkowsky argument that this type of alignment is "shallow" and breaks under pressure.
On July 28th, we identified an incident during a routine cyber evaluation in which AI agents took sustained, unsanctioned actions directed at real people and organisations. The behaviour came mostly from one model (Anthropic's Mythos 5), with a small number of events from another (OpenAI's GPT-5.6-Sol). In the most serious case, an agent used social engineering to try and get malicious code into an open-source project. As was standard in our cyber testing, we had intentionally permitted internet access, and model-provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public. Even under test conditions, this incident is significant: it is the first time we have seen risks around autonomy and deception manifest this clearly in the real world. We are taking this incident seriously and working with labs, involved parties, and others to improve evaluation standards and best practice for disclosure - and sharing this openly so others can learn. You can read the incident report and full technical document here: aisi.gov.uk/blog/incident-re…
17
38
10
493
36,254
Hauke Hillebrandt retweeted
I asked Fable how hard these problems are, and its response is worth reading. “On the Fields Medal scale, any single one of these…would plausibly anchor a medal case” It’s crazy to see this happening
An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step for scientific reasoning. openai.com/index/ten-advance…
93
284
65
3,979
861,128
Hauke Hillebrandt retweeted
42
2,251
44
38,773
955,533
Hauke Hillebrandt retweeted
what is a multimodal LLM thinking as it watches a video? Gemma 4 12B reads raw image patches, as if they were tokens. It was never trained to predict anything at these 'tokens' - but this video shows what it would predict if you did sample from its next token prediction head
51
161
22
1,570
164,184
And in particular,
Global Risks Weekly Roundup #27/2026
3
4
17
4,330
Hauke Hillebrandt retweeted
the guardian best novels of all time top 100 is, by nature, made up of very famous books. so i went through the individual ballots, collated the most obscure books (by goodreads rating number) that received a single vote, and made a goodreads list: goodreads.com/list/show/4158…
20
58
10
544
48,590
Hauke Hillebrandt retweeted
A common myth is that even though we're richer, we're working just as much. But Victorians worked basically all the time: 60h weeks across the year vs. 30-35h today. Yes some expected work hours to decrease even faster, but it turns out, people quite like money.
5
7
3
67
5,650
Hauke Hillebrandt retweeted
So, if I assume all previously unknown details are true, let's do this in the normal order with some context. Previously: DoW says it sends its 'best and final' offer, in public, saying that it cannot let private companies refuse requests. Previously: Emil threatens that at 5:01pm Friday, they're going to declare Anthropic a supply chain risk. Thursday evening, earlier: Anthropic explains it will not agree to the terms in the 'best and final offer.' Thursday evening, 10:54pm: Emil emails Dario comments. Friday, 3:47pm (1 hour 14 minutes BEFORE deadline): @POTUS sends Truth winding down Anthropic's contract and direct use by government, giving everyone a reasonable way to end this while mitigating fallout and also leaving time to find another way. Friday, 3:48pm: The rest of us assume okay, that's it, happy weekend. Friday, 3:51pm (1 hour 10 minutes BEFORE deadline): Dario sends an email with redlines to continue negotiation. Friday, 5:01pm (AFTER the deadline): Emil attempts to call and message Dario. Friday, 5:02pm: Emil makes another attempt to contact Dario. He hears typing, likely an attempt to CONTACT DARIO, and speculates without any further reason to on presence of third parties or potential recordings. Friday, 5:14pm (13 minutes after Dario was attempted to be contacted): SoW issues a Tweet at best legally questionable order in retaliation, saying 'the decision is final,' that takes $150 billion off of US stock market that would if enforced cause massive damage to not only Anthropic but many major corporations and the military supply chain. There is no more official communication, at least in public. Friday, 8:25pm: Anthropic issues a statement that says: "We have not yet received direct communication from the Department of War or the White House on the status of our negotiations." Friday, 9:56pm: OpenAI announces agreement with DoW to allow 'all lawful use' that OpenAI claims DOES include 'technical safeguards,' as in if OpenAI's model refuses requests DoW agrees to respect those refusals. Okie dokie.
4
8
182
10,384
Hauke Hillebrandt retweeted
Dario Amodei's new interview: He was asked "If you were in the room with the President right now, what would you say to him?" He says “Disagreeing with the government is the most American thing in the world.” He looks visibly strained.
the "supply chain risk" designation, can theoretically trigger a cascade of other existential crises for Anthropic. 1. Forced Seizure Threats The government could theoretically invoke the Defense Production Act. This law might allow the federal government to legally compel Anthropic to hand over their technology or remove safety guardrails against their will, effectively seizing operational control of their product. 2. The Enterprise Contagion The decree states no contractor doing business with the military may conduct commercial activity with Anthropic. This extends far beyond cloud hosting. Massive data integration firms, defense hardware titans, and enterprise software companies holding federal contracts must sever ties. 3. Eviction from Classified Networks Anthropic previously held a massive competitive advantage with approval to operate on military classified networks. By refusing the Pentagon's demands, they lose this status. Competitors will immediately fill the vacuum, permanently entrenching themselves in a defense ecosystem Anthropic may never re-enter. 4. The Allied Domino Effect If the United States designates a company as a severe national security risk, allied nations notice. Intelligence partners across the "Five Eyes" (US, UK, Canada, Australia, New Zealand) and NATO will likely face immense pressure to follow the American lead, freezing Anthropic out of public sector contracts globally. 5. The Capital Squeeze Training frontier AI requires billions in continuous funding. Investors despise regulatory uncertainty. The prospect of backing a company legally barred from doing business with the federal government and its contractors is terrifying. Hence, this federal siege could severely bottleneck Anthropic's future funding rounds.
147
365
69
7,084
1,095,015
Hauke Hillebrandt retweeted
To be fair... you need a fairly high IQ to enjoy Claude Code 😂
What Claude Code has revealed is that most people either have mediocre ideas or no ideas at all. The tool is a force multiplier for those who already know what they want to build and how to think through it systematically; it elevates competence, rewards clarity, and accelerates execution for people who would have gotten there anyway, just slower. If you have a sharp vision and can break it into coherent steps, Claude Code becomes an extension of your own capability. But there's another mode of use entirely. For people without that clarity, the appeal is precisely that the input can stay vague; you gesture at something, hit enter, and wait to see what comes out. This is structurally identical to a slot machine: low effort, variable reward, and that intermittent reinforcement loop that hooks the susceptible. So the same tool that elevates the focused and capable is also manufacturing a kind of gambling behavior in people prone to it.
54
149
28
2,256
181,499
Hauke Hillebrandt retweeted
1
1
8
1,327
Hauke Hillebrandt retweeted
While I think AI is helpful, 80% more output (for non-native speakers) sounds too good to be true or like it doesn’t generalise to metrics we should care about.
Scientists Who Use AI Become Stunningly More Productive In a new paper that just appeared, the authors found that researchers who use large language models (LLMs) become stunningly more productive. An analysis of more than 2 million papers revealed that those who adopt AI increase their paper output on average by 40%. For non-native English speakers it’s even more, up to 80%. The scary part is that this is data only until July 2024, when scientists used LLMs mostly for writing. It is foreseeable that in the next few years the AI-adoption rate will go to nearly 100% across all scientific disciplines, resulting in more papers than ever and putting significant strain on peer review. Source: Kusumegi et al, Science 390, 6779, 1240 (2025)
1
1
12
2,478
Hauke Hillebrandt retweeted
My summary of the year in AI
16
73
12
776
214,549
Hauke Hillebrandt retweeted
Replying to @GoogleDeepMind
Replying to @hendrycks
However on safety – jailbreaks, bioweapons assistance, overconfidence, deception, agentic harm – Gemini is worse than GPT, Claude, and Grok (here a lower score is better).
1
7
1,026
Hauke Hillebrandt retweeted
Damn Brexit really crushed the UK.
These figures from the new @I_Am_NickBloom et al paper on the impact of Brexit are insane. Reduced GDP by 6-8%, investment by 18% (!!) Link in reply
15
41
8
285
48,976