Embed Notice

HTML Code

<blockquote style="position: relative; padding-left: 55px;"><section><a href="https://someone.elses.computer/users/mikarv/statuses/112755858406569264">Michael Veale (mikarv@someone.elses.computer)'s status on Tuesday, 09-Jul-2024 21:07:47 JST</a><a href="https://someone.elses.computer/@mikarv" title="mikarv@someone.elses.computer"><img src="https://gnusocial.jp/avatar/14401-48-20230911222354.webp" width="48" height="48" alt="Michael Veale" style="position: absolute; left: 0; top: 0;">Michael Veale</a><div><ul><li></ul></div></section><article><p>The new Tony Blair Institute report on public sector automation savings classifies the potential to automate a task based on… how much GPT-4 classifies a task as automatable. </p><p>Is this the worst method ever? It’s absurd, unserious, and any outlet reporting on this uncritically should be humiliated and ridiculed. (seen via <a href="https://mastodon.social/@halcyene">@halcyene</a>)</p></article><footer><a rel="bookmark" href="https://gnusocial.jp/conversation/3350722#notice-6596550">In conversation</a><time datetime="2024-07-09T21:07:47+09:00" title="Tuesday, 09-Jul-2024 21:07:47 JST">about 10 months ago</time> <span>from <span><a href="https://someone.elses.computer/@mikarv/112755858406569264" rel="external" title="Sent from someone.elses.computer via ActivityPub">someone.elses.computer</a></span></span><a href="https://someone.elses.computer/@mikarv/112755858406569264">permalink</a><h4>Attachments</h4><ol><li><label><a rel="external" href="https://gnusocial.jp/attachment/2885432">We therefore use a hybrid of the two approaches. We first use GPT-4
to categorise each of the 19,281 tasks in the O*NET database in several
different respects that we consider to be important determinants of whether
the task can be performed by Al or not. These were chosen following an
initial analysis of GPT-4's unguided assessment of the automatability of
some sample tasks, in which it struggled with some assessments. This
categorisation enables us to generate a prompt to GPT-4 that contains an
initial assessment as to whether it is likely that the task can or cannot be
performed by Al. In the final stage, we ask GPT-4 to categorise the type
of Al tool that could be used to perform the task, estimate the time saving
and give an assessment of whether it would be cost effective to deploy the
tool. We then merge this information with UK Labour Force Survey data to
calculate overall time savings given the numbers in each occupation group.</a></label><br><a href="https://inside.someone.elses.computer/computer/media_attachments/files/112/755/857/902/712/145/original/9a90e387f8061848.png" rel="external">https://inside.someone.elses.computer/computer/media_attachments/files/112/755/857/902/712/145/original/9a90e387f8061848.png</a></li><li><label><a rel="external" href="https://gnusocial.jp/attachment/2885433">We find that around 40 per cent of tasks performed by public-sector
workers could be done by Al. However, relatively few of these tasks could
be fully replaced by Al, meaning some worker involvement would still be
necessary in almost all cases. Taking our estimated time savings from using
these tools into account, our analysis suggests that 19.8 per cent of public-
sector workers' time could be saved using Al tools.</a></label><br><a href="https://inside.someone.elses.computer/computer/media_attachments/files/112/755/858/160/537/896/original/e8db705284cfc6f3.png" rel="external">https://inside.someone.elses.computer/computer/media_attachments/files/112/755/858/160/537/896/original/e8db705284cfc6f3.png</a></li></ol></footer></blockquote>

Corresponding Notice

Embed this notice
Michael Veale (mikarv@someone.elses.computer)'s status on Tuesday, 09-Jul-2024 21:07:47 JSTMichael Veale
- Matt Davies
The new Tony Blair Institute report on public sector automation savings classifies the potential to automate a task based on… how much GPT-4 classifies a task as automatable.
Is this the worst method ever? It’s absurd, unserious, and any outlet reporting on this uncritically should be humiliated and ridiculed. (seen via @halcyene)
In conversationabout 10 months ago from someone.elses.computerpermalink
Attachments
1. We therefore use a hybrid of the two approaches. We first use GPT-4 to categorise each of the 19,281 tasks in the O*NET database in several different respects that we consider to be important determinants of whether the task can be performed by Al or not. These were chosen following an initial analysis of GPT-4's unguided assessment of the automatability of some sample tasks, in which it struggled with some assessments. This categorisation enables us to generate a prompt to GPT-4 that contains an initial assessment as to whether it is likely that the task can or cannot be performed by Al. In the final stage, we ask GPT-4 to categorise the type of Al tool that could be used to perform the task, estimate the time saving and give an assessment of whether it would be cost effective to deploy the tool. We then merge this information with UK Labour Force Survey data to calculate overall time savings given the numbers in each occupation group.
  https://inside.someone.elses.computer/computer/media_attachments/files/112/755/857/902/712/145/original/9a90e387f8061848.png
2. We find that around 40 per cent of tasks performed by public-sector workers could be done by Al. However, relatively few of these tasks could be fully replaced by Al, meaning some worker involvement would still be necessary in almost all cases. Taking our estimated time savings from using these tools into account, our analysis suggests that 19.8 per cent of public- sector workers' time could be saved using Al tools.
  https://inside.someone.elses.computer/computer/media_attachments/files/112/755/858/160/537/896/original/e8db705284cfc6f3.png

Public

Embed Notice

HTML Code

Corresponding Notice