Fergus Henderson

[Demis Hassabis @demishassabis]

⚡Iterating at lightning speed! Gemini 4 Carbon is SOTA across key benchmarks and combines frontier capabilities with deep safety & security research.

[Fergus Henderson @fergushenderson]

The floodwaters are rising...

[Fergus Henderson @fergushenderson]

@JonathanRoss321 Wow, Groq is really taking off!

[Fergus Henderson @fergushenderson]

@QAI_Dev @jackclarkSF @OpenAI There will be no difference between "successfully simulating AGI" and AGI.

[Fergus Henderson @fergushenderson]

@jtwagnernebras1 @PicoPaco17 @tegmark @ylecun, or other skeptics of imminent AGI/ASI arrival, could post a cryptographic hash (e.g. SHA-512) of a task now, and then post the actual task in 2027.

[Fergus Henderson @fergushenderson]

@adamscochran @conatux The use of "we" in an Abstract is just the standard academic "we", and does not necessarily indicate that there was more than one author of the paper. This is not useful evidence for Satoshi hunters.

[Fergus Henderson @fergushenderson]

@lynetcha @AndrewYNg Copyright protects the right to create derivative works. Using an LLM to rephrase a copyrighted work in different words is creating a derivative work, and is infringement — unless fair use. Doing this in bulk, without adding additional value, for commercial gain? Not fair use.

[Fergus Henderson @fergushenderson]

@satnam6502 O.M.G.!!!

[Fergus Henderson @fergushenderson]

@NecroKuma3 @GaryMarcus Waymo is an Alphabet company that started as the Google self-driving car project. "Waymo One ride-hailing service operates in Metro Phoenix, San Francisco, and is ramping up in Los Angeles County and Austin, Texas. We look forward to expanding further."

[Fergus Henderson @fergushenderson]

@halvarflake 6) Yes, there's incentive to self-deceive about the importance of the risks. But we're not wrong about the importance and impact of AI and AGI in the long term, and the risks are real, including the risk that we might only get one chance to get it right. So caution seems wise.

[Fergus Henderson @fergushenderson]

@halvarflake 5) It is possible to do plenty of harm just by interacting with the world via the internet.

[Fergus Henderson @fergushenderson]

@halvarflake 4) Orcas can't communicate in human languages or connect to the internet.

[Fergus Henderson @fergushenderson]

@halvarflake 3) We have seen & continue to see exponential increases in raw computational power, and there's reason to believe AGI capabilities will scale with that. Even if there's no rapid take-off, capture of economic power may be quite fast once a certain threshold is reached.

[Fergus Henderson @fergushenderson]

@halvarflake 2) A human of vastly higher IQ than average cannot easily or quickly clone themself or increase the size of their brain. An ASI could clone themself much more quickly than a human could, and could increase the size of their own computational substrate as much as resources allow.

[Fergus Henderson @fergushenderson]

@halvarflake I am not an "AI doomer". But:
1) It seems likely that an ASI merely thinking and interacting on the internet can (a) generate revenue, (b) rent compute power, (c) solve robotics, (d) use robots to do whatever experimentation is needed for other engineering problems.

[Fergus Henderson @fergushenderson]

@svpino Humans have excellent generalisation skills, and mediocre memorization, compared to machines. If a human can pass a test, it is very likely that they will be able to answer similar questions. For an LLM, not necessarily. The problem of tests in the training data is real.

[Fergus Henderson @fergushenderson]

@AstroMikeMerri @ProfBrianCox Many Worlds QM interpretation avoids the fine-tuning problem. Even if "none" is much more likely than "one", there can still be many universes with one civilization.

[Fergus Henderson @fergushenderson]

@rodneyabrooks Exponential reductions in the cost of compute ($/op) are not over yet, not by a long way.

[Fergus Henderson @fergushenderson]

@brendan_evers @sytelus @ashertrockman That's a footnote, and it applies only to the first data point, not to the other data points. "excludes imports" in this context means not counting the characters in the "import ..." lines. Obviously none of these include all the lines of the library modules, language impl. & OS.