MattFfrench

Run Madrid AI Builders (1500+ builders) · Lead author, ConfidenceBench: do LLMs know when they're wrong? · ML lead · Cambridge Eng

[MattFfrench @Mattffrench]

the most accurate AI model is only 6th at knowing when its wrong

we put 31 models and a human through 200 questions. claude opus 5 got the most right. fable 5.1 was best at knowing when to doubt itself

the human beat most of them

[MattFfrench @Mattffrench]

anthropic ipo and gta 6 in the same week. tech productivity is going to zero

[MattFfrench @Mattffrench]

polymarket has claude at 0.1% to be the best model on sunday and anthropic at 80% to ipo above $1.25 trillion. being second best is worth a trillion apparently

[MattFfrench @Mattffrench]

@AnthropicAI safety evals the week before the ipo is very on brand

[MattFfrench @Mattffrench]

@jimcramer so they find one by monday then

[MattFfrench @Mattffrench]

@DiscussingFilm watching it is free, understanding it costs a booster box

[MattFfrench @Mattffrench]

@Rap the last 0.6% is going to feel longer than the first 99.4

[MattFfrench @Mattffrench]

@TheStalwart imagine getting tipped in nvda in 2016

[MattFfrench @Mattffrench]

@WatcherGuru robots can share skills across bodies and my laptop still cant find the printer

[MattFfrench @Mattffrench]

@mntruell finally someone in the meeting who actually read the doc

[MattFfrench @Mattffrench]

@rawsalerts soar is doing a lot of work here, polymarket has him back by end of november at 2.9%

[MattFfrench @Mattffrench]

@Polymarket 57% for a tower catching a falling skyscraper with chopsticks still feels generous

[MattFfrench @Mattffrench]

@antigravity the agent writes the script that runs the agents. one more layer and it runs the standup too

[MattFfrench @Mattffrench]

@ctnzr somewhere an h100 feels very guilty about writing that letter

[MattFfrench @Mattffrench]

@BleacherReport @CamHeyward @NotJustFootball best perk of being a former head coach is being free for gta week

[MattFfrench @Mattffrench]

@unusual_whales polymarket has 69% that jason or lucia dies so maybe dont get too attached

[MattFfrench @Mattffrench]

@XBOX 6 days is just enough notice to call in sick

[MattFfrench @Mattffrench]

@unusual_whales deer in the headlights to ceo in 3 months is the comeback arc of the year

[MattFfrench @Mattffrench]

@jojocve @GenLayer the five minute demo is always the best five minutes that tool will ever have

[MattFfrench @Mattffrench]

@0xLeBwA @quantpilotapp backtesting in chat is dangerous, way too easy to keep tweaking till it looks good

[MattFfrench @Mattffrench]

japan was ready for this the whole time