SBIRs

• Demos
• More paper

GPT Agents

• Store more NZ twitter in db
• Found this: Social Dominance: An Intergroup Theory of Social Hierarchy and Oppression

aitextgen is a Python package that leverages PyTorch, Hugging Face Transformers and pytorch-lightning with specific optimizations for text generation using GPT-2, plus many added features. It is the successor to textgenrnn and gpt-2-simple, taking the best of both packages (Github repo)

I think this might be a way to train a GPT model from scratch that is trained in reverse. Which means, given an outcome, it should be possible to create a cluster of text options that would construct that string on a regular GPT. And it should be possible to verify by using that phrase on a forward GPT to evaluate the likelihood that the originating/target text gets generated.

Doctor!

SBIRs

• Continuing to work on paper. I want to add a section that talks about how our biases affect our tactical and strategic decisions. From Ethics Education for Irregular Warfare:
  • For the four members of SEAL Team 10, inserted in the Hindu Kush mountains of Afghanistan’s Kunar province on the night of June 27, 2005, however, the problem was not this puzzling abundance of private contractors and security firms in the battlespace, but a decided absence of reinforcements or backup support of any kind in a remote and inaccessible region far from their operational headquarters. Codenamed ‘Operation Redwing’, the mission of these Special Forces personnel was to reconnoiter and get ‘eyes on’ Ahmad Shah, a close associate of Osama bin Laden, whose attacks had been taking a heavy toll on Marines operating in eastern Afghanistan. After setting up their observation post on a mountainside, overlooking a village near the Pakistani border in which this key Taliban leader was believed to be encamped with a small army, the four-man team was approached at midday by two Afghan men and a 14-year-old boy, herding their flock of goats. The SEALs debated over whether to kill the three civilians in order to protect their cover, try to hold them prisoner, or simply turn them loose and abandon the mission. After arguing among themselves, the four SEALS decided to let the Afghans go, and attempt to re-position. A little later, however, nearly one hundred Taliban fighters materialized, coming across the same ridge over which the goat-herds themselves had fled. The SEAL team fought for several hours, killing an estimated 35 of the enemy, but eventually they were overwhelmed. Their commanding officer, US Navy Lieutenant Michael Murphy, was shot and killed as he called for backup.

• In this case, Lt. Murphy made the ethical choice, and the proximal cause of his death was that choice. However, the lack of support plays into the result as well, and at a still higher level, the abandonment of Afghanistan to the Taliban shows that many of the decisions made in that 20-year campaign were deeply flawed.
• We, as human beings have many biases. These may involve gender, ethnicity, and race. But we also have less obvious biases also affect how we make decisions on those issues that affect us such as national security. For example, the USA (among other countries) has a bias towards advanced weapons systems [citation needed]. This is reflected in the decisions to incorporate AI/ML into the nations military. The focus is on intelligent munitions, drones, hypersonic missiles, etc. But since the end of the cold war (in 19xx?), the majority of military operations have been in irregular conflict, such as Kosovo, Somalia, and Afghanistan. An intelligent munition would not have helped Lt. Murphy’s team decide whether to kill, hold, or release the Afghan shepherds that stumbled upon them. But information presented in a way that lets a user clearly visualize the likely outcome of a trajectory of choices, may well choose another path. After Vietnam, Iraq, and Afghanistan, leaders might think twice if that they see they are heading towards the part of the map marked “Quagmire”.
• Done with this pass!

SBIRs

• Working on paper. Currently listening to George Lucas (Not that George Lucas) talking about military ethics

• This also looks interesting: On Obedience
  • an in-depth and nuanced philosophical treatment of the virtue of obedience in the context of the professional military and the broader civilian political community, including the general citizenry. The nature and components of obedience are critical factors leading to further discussions of the moral obligations related to obedience, as well as the related practical issues and implications. Pauline Shanks Kaurin seeks to address the following questions: What is obedience? Is it a virtue, and if it is, why? What are the moral grounds of obedience? Why ought military members and citizens be obedient? Are there times that one ought not be obedient? Why? How should we think about obedience in contemporary political communities?
• Human Terrain System: was a United States Army, Training and Doctrine Command (TRADOC) support program employing personnel from the social science disciplines – such as anthropology, sociology, political science, regional studies, and linguistics – to provide military commanders and staff with an understanding of the local population (i.e. the “human terrain”) in the regions in which they are deployed.
  • The concept of HTS was first developed in a paper by Montgomery McFate and Andrea Jackson in 2005,^[6] which proposed a pilot version of the project as a response to “identified gaps in [US military] commanders’ and staffs’ understanding of the local population and culture”, such as became particularly visible during the US invasion of Iraq and Afghanistan.^[1][3][4] HTS was subsequently launched as a proof-of-concept program, run by the United States Army Training and Doctrine Command (TRADOC), in February 2007, with five HTS teams deployed between Iraq and Afghanistan.^[3][4] Since 2007, HTS has grown from a program with five deployed teams and a $20 million two-year budget to one with 31 deployed teams and a $150 million annual budget.^[3] HTS became a permanent US Army program in 2010.
  • Ever since its launch, HTS has been surrounded by controversy.^[4][7][8] While the program initially received positive coverage in the US media, it quickly became the subject of heavy criticism – particularly from anthropologists, but also from journalists, military officials and HTS personnel and former personnel.^[9] Most notably, on 31 October 2007, the Executive Board of the American Anthropological Association (AAA) published a statement opposing HTS as an “unacceptable application of anthropological expertise” that conflicted with the AAA’s Code of Ethics.^[10][11][12] Following the publication of a report on HTS by the Commission on Engagement of Anthropology with the US Security and Intelligence Services (CEAUSSIC) in 2009,^[13][14] the AAA released a further statement of disapproval, which they re-iterated in 2012 after rumours that the controversy had died down.

• Got together with Aaron for a couple of hours of writing

Called principal!

Need to drop off NASA badge

GPT Agents

• The gml export rewrite yesterday worked! Here’s the same data from yesterday with the source and target nodes:

• I think I may try separate source and target files so that I can color links?
• Twitter API and Python stuff:
  • An Extensive Guide to collecting tweets from Twitter API v2 for academic research using Python 3
  • Twitter’s API Page
  • tweepy Twitter for Python
  • twarc a command line tool and Python library for collecting JSON data via the Twitter API, with a command (twarc2) for working with the v2 API
  • python-twitter a simple Python wrapper for Twitter API v2
  • TwitterAPI minimal Python wrapper for Twitter’s APIs
  • twitterati Wrapper for Twitter Developer API V2
  • twitter-stream.py a Python API client for Twitter API v2
  • twitivity Account Activity API client library for Python
  • PyTweet a synchronous Python wrapper for the Twitter API
  • tweetkit a Python Client for the Twitter API for Academic Research
  • 2wttr get Tweets from the v2 Twitter API, for Academic Research
• 4:15 meeting

SBIRs

• Continuing on paper. Down to 12.5 pages! Page limit s not hard. More ethics discussion is preferred

Today I learned that I’m working on Computational Sociolinguistics

GPT Agents

• Found a bug with my GML code. I seem to be saving only the source->target relationships

• Need to explicitly add the target->source info as well, but I think there are multiple maps. The image above is the outbound map. It might make sense to have inbound and both
• Rewrote the entire GML export to use database queries. Makes more sense now

SBIRs

• 9:15 standup
• Working on the paper. Now at 13.5 pages. Working on Results

Having a hard time getting started today

Try calling Principal again

SBIRs

• 10:00 Belief map meeting. Just discussed getting the paper through the review process
• Working on paper – finished cleaning up the intro and am working on background

GPT Agents

• Good meeting with Andreea. She introduced me to sociolinguistics, which looks to be the root of a lot of the ‘unique populations’ that we can train a GPT to mimic. William Labov (en.wikipedia.org/wiki/William_Labov) is one of the founders in the field. Got some of his books and Introducing Sociolinguistics by Miriam Meyerhoff (scholar)

GPT Agents

• Got my Twitter developer account!
  • I need to see if I can figure out threading. This looks like it shows how to do it: developer.twitter.com/en/docs/twitter-api/conversation-id
  • Need to get the full academic license – done!
• Test topic connections
  • Write and read to db – maybe? Verify. Seems to be working!
  • Continue with conspiracy theories
• Had a some interesting ideas.
  • Once a map is built, narratives can be created by the GPT, either in standalone or in dialog with others. The trajectory of these stories across the map to see the major flows. In turn, the flows can be used to adjust layout.
  • The re-clustering can still use the group labels as the ‘topics’. So we might get Jews/Puppetmasters and Government/False Flags. The order can depend on the ratio
• GML output
• Got the groups writing out:

And the topics in the groups. Here’s PuppetMasters:

SBIRs

• GML output
• Work with Aaron on the RFI text?
• Work the paper (due Dec 10).
  • Single-column IEEE format – done and on Overleaf using their template
  • 12 pages (Currently at 15!)

Get bike! Nope, still needs bearing races

GPT Agents

• Got my Twitter developer account! Now I need to see if I can figure out threading
• Test topic connections
  • Add connections to MapTopic.to_string() – done
  • Pull found text from raw – done
  • Write and read to db – maybe? Verify
  • Continue with conspiracy theories

SBIRs

• Write stories for today’s sprint planning
  • GML output
  • Topic work
  • Refactoring
  • Change player so that it can handle any number of ‘players’
  • Add GML input to player
  • Meeting about tasking and IP. I need to write up something about when patents are good (slowly changing environments, where defense is important) and bad (dynamic, disruptive environments where exploration is easy). Incidentally, the Patent Office needs something like ArXiv, where ideas can be published as prior art so that they cannot be taken from the public domain. It should be a free service.
    • Preliminary meeting with Aaron. Moved to Monday at 10:00
  • The only real task is the paper. Which is due on Dec 10, so that’s probably a good thing

It seems as though I’m kinda burnt out from the last effort

Get bike!

GPT Agents

• Playing around with the conspiracy theory map since I know that area and it needs another pass
  • I have a thought about counting new results from the gpt. When there is a response from the gpt, there should be a button(?) that counts and deletes all substrings in the return that matches a topic in a node, with the appropriate connections between topics and nodes. I’ll need to add a count value to table_topic. It can have a default of 1
• Start generating gml files
• 3:30 Meeting with Andreea. Looked at the word frequencies for the yelp vegetarian data to see if this is the kind of frequencies she mentioned last week as a linguist
• Got my Twitter dev API!

SBIRs

• Sprint demos! Done!
• Add brownbag of technique for next sprint
• Set up meeting to discuss next steps for 4:00 tomorrow
• Got pulled into some RFI thing?

NLP+CSS 201: Beyond the basics

• This website hosts the upcoming tutorial series for advanced NLP methods, for computational social science scholars.
• Every few weeks, we will host some experts in the field of computational social science to present a new method in NLP, and to lead participants in an interactive exploration of the method with code and sample text data. If you are a graduate student or researcher who has some introductory knowledge of NLP (e.g. has learned text analysis from SICSS) and wants to “level up”, come join us!
• Watch past tutorials on our YouTube channel.

OpenAI makes GPT-3 generally available through its API | VentureBeat

Book

• Submitted proposal to Oxford press!

GPT Agents

• Adding connections between topics – done!
• Playing around with the conspiracy theory map since I know that area and it needs another pass
  • I have a thought about counting new results from the gpt. When there is a response from the gpt, there should be a button(?) that counts and deletes all substrings in the return that matches a topic in a node, with the appropriate connections between topics and nodes. I’ll need to add a count value to table_topic. It can have a default of 1
• Monday start generating gml files

SBIRs

• LAIC Demo! Then go for a long ride
• Demo is DONE! I think it went well

GPT Agents

• Add in topic links
• Add check Wikipedia button? I guess at the bottom row of buttons – done
• Work on GML output once the topic links are in
• Start conspiracy map

GPT-Agents

• Add correlation to the columns of the LIWC data – done
• When saving as GML, export the MapGroup as one file and then each set of MapTopics for that group as a file – started
• 4:15 Meeting – went over the paper and the new spreadsheets

SBIRs

• Add ground-truth checking to slides – done
• Add qualitative slide – done
• Added some screenshots for the live demo
• Walk through presentation with Aaron. Done.
• Change the script to match the current. Done
• Add a final slide?

Book

• Add Resume/CV

JuryRoom

• 6:00 Meeting. Lots of work on Jarod’s stuff. Did a practice walkthrough of the LAIC work

Registered for graduation! Done!

Set up physical appt

JuryRoom

• Continuing to run Andreea’s probes
• Responded to Jarod’s email on large transformer language models
• Meeting with Andreea today

SBIRs

• 9:15 standup – done
• Working on adding connections between topics: TODO: add self.connect_topics(self.selected_topic, self.selected_seed_topic) # Verify this works!
  • Added connection_set:Set to MapTopic, but no implementation yet

GPT Agents

• Send draft to Jimmy and Shimei – done

Book

• Evaluate proposal and send in?

SBIRs

• Worked most of the weekend. I have a first pass on the methods and results section, some notes on discussion, and still need to start the conclusions. Also, fold the text into the white paper – DONE!
• Meetings with Aaron?

GPT Agents

• Got LIWC data from Shimei
• Need to run probes for Andreea – running