Embed Notice

HTML Code

<blockquote style="position: relative; padding-left: 55px;"><section><a href="https://dair-community.social/users/emilymbender/statuses/111802010065000023">Prof. Emily M. Bender(she/her) (emilymbender@dair-community.social)'s status on Tuesday, 23-Jan-2024 23:17:37 JST</a><a href="https://dair-community.social/@emilymbender" title="emilymbender@dair-community.social"><img src="https://gnusocial.jp/avatar/76831-48-20221219012403.webp" width="48" height="48" alt="Prof. Emily M. Bender(she/her)" style="position: absolute; left: 0; top: 0;">Prof. Emily M. Bender(she/her)</a></section><article><p>I feel very vindicated for not making time to  answer journalist's queries about papers that "prove" things based on hypothesized graphs and fabricated data.</p></article><footer><a rel="bookmark" href="https://gnusocial.jp/conversation/2624874#notice-5199048">In conversation</a><time datetime="2024-01-23T23:17:37+09:00" title="Tuesday, 23-Jan-2024 23:17:37 JST">Tuesday, 23-Jan-2024 23:17:37 JST</time> <span>from <span><a href="https://dair-community.social/@emilymbender/111802010065000023" rel="external" title="Sent from dair-community.social via ActivityPub">dair-community.social</a></span></span><a href="https://dair-community.social/@emilymbender/111802010065000023">permalink</a><h4>Attachments</h4><ol><li><label><a rel="external" href="https://gnusocial.jp/attachment/2153447">Screenshot: The researchers started by assuming that there exists a hypothetical bipartite graph that corresponds to an LLM’s behavior on test data. To explain the change in the LLM’s loss on test data, they imagined a way to use the graph to describe how the LLM gains skills.Take, for instance, the skill “understands irony.” This idea is represented with a skill node, so the researchers look to see what text nodes this skill node connects to. If almost all of these connected text nodes are successful — meaning that the LLM’s predictions on the text represented by these nodes are highly accurate — then the LLM is competent in this particular skill. But if more than a certain fraction of the skill node’s connections go to failed text nodes, then the LLM fails at this skill.Source: https://www.quantamagazine.org/new-theory-suggests-chatbots-can-understand-text-20240122/</a></label><br><a href="https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/802/007/667/604/512/original/8f750a2c60fd6162.png" rel="external">https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/802/007/667/604/512/original/8f750a2c60fd6162.png</a></li><li><label><a rel="external" href="https://gnusocial.jp/attachment/2153448">Screenshot: The team also automated the process by getting GPT-4 to evaluate its own output, along with that of other LLMs. Arora said it’s fair for the model to evaluate itself because it doesn’t have memory, so it doesn’t remember that it was asked to generate the very text it’s being asked to evaluate. Yasaman Bahri, a researcher at Google DeepMind who works on foundations of AI, finds the automated approach “very simple and elegant.”Source: https://www.quantamagazine.org/new-theory-suggests-chatbots-can-understand-text-20240122/</a></label><br><a href="https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/802/008/580/241/781/original/4cf2efa360e2644e.png" rel="external">https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/802/008/580/241/781/original/4cf2efa360e2644e.png</a></li><li><label><a rel="external" href="https://gnusocial.jp/attachment/2153449">Screenshot: Nonetheless, Hinton thinks the work lays to rest the question of whether LLMs are stochastic parrots. “It is the most rigorous method I have seen for showing that GPT-4 is much more than a mere stochastic parrot,” he said. “They demonstrate convincingly that GPT-4 can generate text that combines skills and topics in ways that almost certainly did not occur in the training data.” (We reached out to Bender for her perspective on the new work, but she declined to comment, citing a lack of time.)Source: https://www.quantamagazine.org/new-theory-suggests-chatbots-can-understand-text-20240122/</a></label><br><a href="https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/802/009/441/472/197/original/da8dfa442b1159a3.png" rel="external">https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/802/009/441/472/197/original/da8dfa442b1159a3.png</a></li></ol></footer></blockquote>

Corresponding Notice

Embed this notice
Prof. Emily M. Bender(she/her) (emilymbender@dair-community.social)'s status on Tuesday, 23-Jan-2024 23:17:37 JST Prof. Emily M. Bender(she/her)
I feel very vindicated for not making time to answer journalist's queries about papers that "prove" things based on hypothesized graphs and fabricated data.
In conversationTuesday, 23-Jan-2024 23:17:37 JST from dair-community.socialpermalink
Attachments

Public

Embed Notice

HTML Code

Corresponding Notice