An overview of the Climate Finance Tracker, an initiative launched with OneEarth. Each circle or node on the map represents one private company in the U.S. Users can sort by themes, sectors, investors, and more.

The emerging threat of climate change to our national security

From as early as 2008, the U.S. Department of Defense flagged the increased risk to national security if climate change wasn’t urgently addressed by the federal government.

“We judge global climate change will have wide-ranging implications for U.S. national security interests over the next 20 years,” wrote the National Intelligence Committee during President George W. Bush’s tenure.  The political will to take action to address climate change as a national security issue is slowly growing. Recently the U.S. Army released its first ever climate strategy, outlining its goals to cut emissions in half by 2030, produce electric combat vehicles, and advance officer training for a warmer climate. 

At Primer, we build AI technology that helps subject matter experts and decision-makers identify, understand, and act on emerging threats. We believe that the intersection between climate change and national security is imperative to proactively address. 

“Climate change will affect the Department of Defense’s ability to defend the nation and poses immediate risks to U.S. national security.”

-U.S. Department of Defense, 2014 Climate Change Adaption Roadmap

“Increasing migration and urbanization of populations are also further straining the capacities of governments around the world and are likely to result in further fracturing of societies, potentially creating breeding grounds for radicalization.”

-U.S. National Intelligence Strategy, 2019

Our partnership with OneEarth.org

On the cusp of Climate Week in New York starting Sept. 19, Primer is proud to announce our partnership with OneEarth.org, and the public release of the Climate Finance Tracker that we worked on together this year. Leveraging Primer’s machine learning technology on publicly available data, the tracker “follows the money,” surfacing in one unified view, where U.S. private funding for combating climate change is going. In short, following the money helps everyone collaborate more effectively to tackle the climate crisis. 

For the first time, the Climate Finance Tracker is live and open-source, and uses AI to advance intelligence and decision cycles, to ultimately improve outcomes on solving a major global challenge.  

The Climate Finance Tracker allows anyone to search the funding landscape of all private companies and nonprofits in the U.S. that are addressing climate-related topics. By surfacing topics from organization and project descriptions provided by Crunchbase and Candid, the tracker allows us to better understand the shape of the current response, and with that, its potential gaps. This understanding means funders can better direct their resources, ensuring the highest possible impact of their investment.

The value of following the funding

Eric Berlow, a leading social impact data scientist and researcher, led the project with his nonprofit Vibrant Data Labs and OneEarth.org. Berlow explained how using Primer’s AI technology allows researchers to find the answer to questions they may not even know to ask. 

“We searched for companies working on known climate-related topics, and then the machine pulled out of their language that some also mention national security-related topics,” Berlow explained. “That’s novel. I wouldn’t have even known to look… And the machine found it.”

Text-based machine learning doesn’t just uncover answers – it also helps surface questions researchers may not even know to ask.

A snapshot of 32 organizations that were tagged with ’national security’, ‘unrest, conflicts and war’, or ‘peace and security.’ Explore the full Climate Finance Tracker here.

“We know climate intersects with national security, and now we can easily see about 25 private investors funding companies in that space,” Berlow explained. “It’s only 6 companies out of 6,000. That’s an important gap. Let’s talk to them and see what’s going on.”

Climate week begins

With this week’s launch of the Climate Finance Tracker, along with Climate Week in New York, we expect to see increased conversation around adoption of machine learning solutions to help solve global problems such as the climate crisis, making us more resilient toward emerging threats.

At Primer, we believe that when we empower the ecosystem with better information, intelligence, and decision making, we can truly solve really complex problems together. 

See the power of text-based AI and NLP. 

Request a demo of Primer’s NLP technology. 

When Russian Military commanders started using unencrypted radio communication, we were able to capture this and deploy our AI tools to extract key insights from the data. What we found were indications of a disorganized army, bogged down, and confused.

Read: “As Russia Plots Its Next Move, an AI Listens to the Chatter,WIRED, April 4, 2022

Troops frequently go out to the battlefield not fully knowing what they might encounter. Even the best armies in the world rarely have all needed capabilities ready and designed for the task at hand. At Primer, we increase operational readiness by building and deploying flexible AI solutions that are optimized for today’s and tomorrow’s defense and intelligence use cases. To help U.S. and strategic partner Services navigate continuously evolving needs, we at Primer believe that the best approach is not a single perfect AI algorithm. The solution is a powerful AI infrastructure that enables subject-matter experts to build, customize, and deploy AI models that help them rapidly process massive volumes of data. When intercepted Russian audio of military commanders discussing tactics recently became available, we saw an opportunity to test our AI platform capabilities. 

Today, linguists and analysts spend their days pouring through noisy radio chatter, on the hunt for actionable information. Analysts are at pains to listen to each minute of audio in hopes of finding useful information and for fear of missing things. Intelligence may only find its way to decision-makers hours later, which could be too late to act on it. Using Primer’s platform capabilities, we can transform these workflows and shorten the time from insight to action. With Primer, users can build and run models across radio chatter that detect Russian language and extract mentions of location, persons, certain topics of interest, or military equipment, summarize text into paragraphs or bullets, and identify trends. By instantly surfacing the audio files and information that is likely most relevant, users cut through the noise, focus on the actual analysis, and develop robust near-real time intelligence for decision-makers. 

How we use AI to transform Russian military audio into operational intelligence 

We started with audio captured from Software Defined Radio (SDR) feeds transmitted over the internet at http://www.websdr.org/ and connected this to our Primer audio ingestion engine which downloaded the communications as wav files. These audio intercepts were then fed into the Primer Platform. First, we used a noise reduction engine and the audio recordings were cleaned to eliminate static or noise, improving sound quality. Cuts containing music or non-Russian speakers were filtered out as they were unlikely to contain our desired targets. These cuts were converted to English text by first using the Russian speech recognition and then Russian to English translation engines within Primer’s platform. Human experts were then able to review the output to understand any areas where the models may have had difficulty.

Sample of the audio captured and then ingested by Primer for analysis

With our cleaned and prepared dataset in hand, we pulled in our commercially available national security models to get a better understanding of what was going down at the frontlines. 

Here are the analytical machine learning capabilities we used to rapidly extract key information from the Russian radio chatter:

  • A topic identification model to find and identify the main audio topics so that analysts can navigate precisely to the audio that they are most interested in.
  • A dialog summarizer to give analysts a concise description and overview of what was said across hours of captured audio.
  • A “call-to-action” model to find and time-stamp orders being issued.
  • A weapons classifier to identify mentions of military equipment.
  • A Named Entity Recognition (NER) engine to identify, extract, and disambiguate all the people, locations, and organizations that are mentioned in the radio communications. 
The AI models from speech-to-text, diarization, and translation showed a pair of soldiers lost and confused.
An AI-generated summary that distills the key points of the dialog saving human analysts critical time.
A NER model extracts people and locations mentioned in the transcript.
Our Military Equipment model identifies mentions of military equipment to help users quickly assess the intelligence value of audio intercepts.
An abstract topic model reduces the main points of an intercept to only a few words, or in this case, user names that are discussed.
The call-to-action model extracts commands from the audio. For the purposes of this blog post, the specific date is obscured to keep exact movements hidden.

Analysts can subscribe to this feed of “calls to action” and consume these in real time on the Primer Platform in apps like Primer Command, or integrate the feed into a Common Operating Picture next to their other battlefield data. This radically democratizes access, as analysts can now directly consume and leverage this information, removing the constraint of needing to be a Russian language specialist or listening for hours to multiple radio frequencies. Instead, analysts can conduct deeper analysis based on a near real time call to action feed of Russian military chatter.

One afternoon’s work on the audio files revealed actionable intelligence that Russian forces were reporting heavy fire in their operating area and losing tanks. They also received orders to retreat from their position. But more importantly we were able to build a custom AI pipeline on our platform to continually and automatically monitor Russian radio chatter and extract tactical insights in near real time. 

This type of real-time intelligence can be a game changer for analysts and commanders. Whoever succeeds in deploying operational AI – reliable AI models that are trained, validated, and in production on live operational data during missions – will have the advantage of faster AI-augmented human decision-making, thereby improving the odds at outmaneuvering the adversary. 

When it comes to battlefield capabilities, the ability to rapidly customize models that meet constantly evolving needs is a crucial differentiator. The models used above are not perfect and will benefit from further training and specialization for specific national security use cases and niches.  Here are a few of the challenges our platform of pre-trained models, labeling, and model training can address to help analysts make immediate workflow improvements in the course of their mission:

  • Our dialogue summarizer model had to contend with the idiosyncrasies of tactical radio communications such as speakers repeatedly saying their own name. In the Primer Platform, this model can be retrained to improve performance, by starting with the machine-generated summary a human analyst can correct and improve. This corrected summary turns into another training datapoint to teach and fine tune the model.  
  • Our weapons classifier model was trained last year during the crisis in the Donbass region of Ukraine. Since then, many more types of military equipment have been deployed. Additionally dialog on radio chatter is much more informal and references to “Helis” are not captured by the initial weapons classifier as “Helicopters.” Users and analyst teams can take the initial models and quickly label new data using our LightTag software and then retrain them to improve their performance in the Primer Platform. 
  • The geolocation engine was not initially trained on tactical radio communication data. As such it was not initially trained to pick up references to map coordinates which are common on radio chatter for artillery units. The users are able to improve the geolocation model by retraining our Named Entity Recognition engine in LightTag and the Platform to identify these and other geographic references.
Creating additional labels within LightTag guarantees the most performant models customized to specific data feeds.

How strategic advantage depends on flexible and customizable AI solutions 

Deployment of AI in a conflict environment is not about having a single perfect algorithm. Instead, the best solution is often a collection of AI algorithms working together as a complex system — the optimal models will change as quickly as the crisis evolves. The most successful approach is to have flexible and customizable AI tools and infrastructure available to users closest to the need so that they can quickly test, customize, and connect the most performant models together. 

At Primer, our goal is to empower those closest to the problem to leverage advanced AI to help them make sense of an incredibly complex information landscape. Machine learning is a force multiplier for linguists and analysts tasked with processing hundreds of hours of audio files.  Small teams can process many hours or days worth of intercepts in a matter of minutes. This enables faster and better tactical responses to unfolding events. Intelligence information previously only available hours or days later can now be in the hands of battlefield commanders when it’s needed most. A strong AI infrastructure for Defense must let users retrain ML models quickly – as these opportunities may not last very long.

Primer’s end-to-end platform provides the AI infrastructure to rapidly build, deploy, and customize models tailored to evolving intelligence and defense use cases.  

The ability to rapidly exploit OSINT, even when messy and in other languages, makes it possible for poorly equipped armies to outmaneuver much better trained and resourced ones. That is a huge paradigm shift, the likes of which we have yet come to terms with. 

Despite having the best trained and resourced services in the world, the hard truth is that a lack of agility and urgency in integrating and operationalizing AI by the U.S. and our strategic allies means that we fall behind faster than we’re willing to admit. As we look ahead, we’re now in a new era of warfare where those with sophisticated and customizable AI capabilities have a distinct advantage in decision superiority, and thereby the ability to outcompete and outmaneuver adversaries to increase the odds of winning wars that cannot be prevented.

For more information about Primer and to access product demos, contact Primer here.

Command’s powerful capabilities can be seen through the lens of Global Security Operation Centers (GSOCs) seeking to provide support and information to employees in Ukraine.

Provide duty of care to employees in Ukraine

For global companies, monitoring for threats to employees or assets requires constant attention. Whether it’s a natural disaster or an act of war, GSOCs are charged with alerting and providing guidance to employees in harm’s way. With Russia’s recent invasion in Ukraine many companies are quickly pivoting to provide support to their employees in the country. Looking back to earlier this year, many people in Ukraine did not think an invasion would happen and are now regretting not evacuating sooner. This is especially true for military aged males who are now being conscripted by the military and unable to evacuate legally. Hearing their employees being separated from their families and in some cases forced to enter military service, employers are expanding their boundaries of care to look after the well-being of their employees and their families in Ukraine as well as Belarus, Russia and the surrounding regions. GSOC’s are being charged with monitoring threats and providing periodic updates to their employees. 

The following showcases how Command helps security teams provide duty of care to their employees with a focus on Kyiv, including:

  • Inform employees about local threats 
  • Update employees on local authority guidance, such as curfew orders
  • Highlight damage to utilities and infrastructure
  • Notify employees about supply shortages

Warn employees about proximity to threats

On February 24th after Russian President Vladimir Putin announced that he had ordered a “special military operation” in eastern Ukraine, missiles began to strike dozens of cities across the country including the capital, Kyiv. By February 27th reports started flooding social media stating the capital was surrounded by Russian forces making evacuations impossible. 

  • Filter by location: The GIF below shows how security teams can quickly filter reporting by their employees location. 
  • Deduplicate: Command automatically collapses all similar reporting so the security analysts do not waste time reading 17 posts about the same Russian convoy marching into Kyiv.
GIF showcasing Command’s filtering and deduplication capabilities.

Update employees on local authority guidance 

As Russian troops continue to advance on Ukraine’s capital, employees in Kyiv need to know the latest guidance from local authorities. 

  • Filter by local authorities: Using the people filter for Kyiv Mayor Vatali Klitschko, social media and news reports related to him and his statements appear. 
  • Key word search: Security teams can further filter by searching for “curfew” to see the latest guidance.
Partial list of people automatically categorized by Command 
Social media post Command tagged as associated with Kyiv Mayor Vatali Klitschko.

Highlight damage to utilities and infrastructure

Employees need to know if there are threats to their power and internet connectivity.

  • Filter by infrastructure: Command is powered by industry leading Natural Language Processing models that accurately identify social posts that fall into this filter. This capability allows security teams to isolate social media reporting related infrastructure and utility damage. 
  • Flag disputed information: While there is a continuous stream of mixed reports about power outages in Kyiv, Command automatically flags disputed information so security teams can accurately track and inform their employees about outages.
Partial list of Command’s filtering options
Examples of Command’s social media results related to internet connectivity. 

Track evacuation options

Some employees in Ukraine are probably seeking support from their employers’ security team on how to evacuate the country safely. 

  • Filter by evacuations: With Command, GSOCs can quickly respond to these requests after narrowing the enormous amount of social media posts related to Ukraine down to just information related to ‘displaced people and evacuations.’ With this filter security teams can summarize social media posts related to routes and border crossing locations. 
  • Map layer: For employees looking to move out of Kyiv to another part of the country, security teams could use Command to filter by ‘Infrastructure Damage’ then select the social mentions map layer to quickly identify areas employees should avoid. 
Partial list of Command’s filtering options and example of Command’s social media results related to evacuations. 

Stay organized and collaborate with team members

Security teams can bookmark and label important reports in their feed. This allows users to stay organized and formulate it into a situation report at a later time. Command’s engineers are also working on the ability to share queries with teammates to form a common operating picture.

Learn more

To watch how Primer Command could help you analyze fast-breaking events, checkout this article posted in AI Business or sign up for a free trial of Command by clicking here. Or contact Primer to discuss your specific needs. You can also stay connected on LinkedIn and Twitter.

Just a few years ago, if you wanted a machine to generate an image for you, you’d need a big data set of the same kind of image. Want to create the image of a human face? You can use a GAN—a generative adversarial network—to generate photorealistic portraits of people who do not exist

But what if you want to generate an image straight from your imagination? You don’t have any data. What you want is to just describe something, in plain language, and have a machine paint the picture.

AI-generated image generated from the prompt: “the relationship between mind and matter”
AI-generated image created from the prompt: “the relationship between mind and matter”

This is now possible. The technique is called prompt-based image generation, or simply “text-to-image”. Its main trick, in a nutshell, is a neural network trained on a vast collection of images with captions. This becomes a Rosetta stone for translating between written language and visual imagery. So when you say “a foggy day in San Francisco” the machine treats your description as a caption and can tell how well any given image fits it. Pair this translator with an image-generating system, such as a GAN, and you have an AI artist ready for your orders.

It was only a matter of months before an online community of AI art hackers emerged. Twitter has become an open gallery of bizarre and wonderful AI-generated images.

Join the revolution

Want to know more? Join me for a panel discussion with four people who are deeply involved with this AI art revolution.

ML Show & Tell: The incredible, bizarre world of AI-generated Art

Now on Demand

Aaron Hertzmann

Aaron Hertzmann is a Principal Scientist at Adobe, Inc., and an Affiliate Professor at University of Washington. He received a BA in Computer Science and Art & Art History from Rice University in 1996, and a PhD in Computer Science from New York University in 2001. He was a professor at the University of Toronto for 10 years, and has worked at Pixar Animation Studios and Microsoft Research. He has published over 100 papers in computer graphics, computer vision, machine learning, robotics, human-computer interaction, perception, and art. He is an ACM Fellow and an IEEE Fellow.

Bokar Lamine N’Diaye

Bokar N’Diaye is a graduate student in the University of Geneva (Switzerland). Currently studying for a master’s degree of anthropology of religions, his interests lie in the intersection of humanities and the digital. Mostly active on Twitter, where he illustrates the computer-generated scriptures of Travis DeShazo (@images_ai), he also collaborates with the Swiss collective infolipo on an upcoming project, Cybhorn, combining music, technology and generative text.

Hannah Johnston

Hannah Johnston is a designer and artist, with an interest in new technologies for creative applications. Following a Bachelor’s in information Technology and a Master’s in Information Systems Science, Hannah spent 9 years designing user experiences at Google. She currently works as a post-secondary instructor and design consultant.

Ryan Murdock

Ryan is a Machine Learning Engineer/Researcher at Adobe with a focus on multimodal image editing. He has been creating generative art using machine learning for years, but is most known for his recent work with CLIP for text-to-image systems. With a Bachelor’s in Psychology from the University of Utah, he is largely self-taught.

Primer partnered with Vibrant Data Labs, a nonprofit that uses data to create the first-ever comprehensive map of climate change funding. Our interview with Eric Berlow of Vibrant Data Labs shows how ‘following the money’ reveals both our biggest opportunities and threats to turn climate change around.

It’s no secret that the future of our planet is in trouble. The recently reported IPCC report concluded that countries aren’t doing nearly enough to protect against significant disasters to come. 

But in order to solve a big problem like climate change, and to understand if our current response is working, we need to see where private funding in the sector is going. What issues are getting money, and which organizations are getting that funding? What other trends might emerge? 

Applying NLP to climate change

That’s where natural language processing comes in. Using Primer’s NLP technology, we partnered with Eric Berlow at Vibrant Data Labs to produce the first-ever climate change funding map. Primer’s Engines analyzed data on over 12,000 companies and nonprofits funded in the last five years. Using organizations’ descriptions provided by Crunchbase, grant applications provided by Candid, and in partnership with the Cisco Foundation, we generated one of the first-ever data-driven hierarchies of climate topics to better understand our current response, alongside any potential gaps. Using this topic hierarchy, we can see what projects organizations are working on – and where. That helps us see what’s missing in the bigger picture. And to solve a problem like climate change, a big picture view is what’s needed.

Watch our interview with Eric Berlow on why following the funding is crucial for the climate’s future. 



“The Coronavirus pandemic was like the trailer to the climate change movie where if the have-nots are really impacted, everybody gets affected by it. Climate change is one of those problems that is an all-hands on deck problem. You cannot just solve a little piece of it. “ – Eric Berlow



Learn more about Primer’s partnership with Vibrant Data Labs here, and learn the technical piece behind the work here and here.

For more information about Primer and to access product demos, contact Primer here.

What is NLP and how can it support business?

Imagine a curious executive walking down a path that explores how Natural Language Processing (NLP) can help their business by uncovering insights hidden in the available and pertinent data to make better and more timely decisions. The executive is often faced with the problem of having vast amounts of company data, and not a lot of ways to take action on it. Not to mention the risk of not knowing what’s hiding in the data.

Steps one and two

As the journey begins, the executive takes steps one and two, Ideate and Identify. The executive asks “What do I want to know?” and follows up with “Where do I find the answers?” Whether it is customer attitudes toward a business, how it compares with competitors, or almost anything else decision-makers, analysts and owners would want to know, a quest for knowledge is the beginning. The next consideration is where to find answers to these questions. Identifying those data sources — internal purchase information, call center logs, product descriptions and reviews, social media posts, customer survey results, etc. — is the “where” that the answers will be found.

Steps three and four

The next steps, three and four, are Connect & Ingest and Transform, where the executive might ask, “How do I find the answers?” extracting text from both external sources and the company’s unstructured internal data mentioned in step two (Identify). In Transform, the executive asks, “How do I use NLP and AI?” focusing on Named Entity Recognition (NER), a sub-task of information extraction. It identifies and classifies named entities mentioned in unstructured data into predefined categories such as given and surnames, affiliated organizations, geographic locations, time expressions, etc. It also includes question & answer, classification, topic modeling, relationship extraction, sentiment analysis and other methods of processing the ingested information into something useful.


Steps five and six

Next up, Integrate and Explore, steps five and six, come after data has been scanned and processed. At the Integrate step, the executive could ask, “How do I combine insights from NLP and AI with my own data and analytical models?” To sharpen the results of NLP, companies often have pre-existing internal mathematical models, analytics and projections that can be combined. Once completed, the executive at the Explore step asks, “What answers do I have?” and looks at the patterns and unearthed relationships that can be converted into action plans.

Steps seven and eight

Operationalize and Realize & Repeat are steps seven and eight. Once the executive has answers from previous steps, the question is “How can I use this information?” The Operationalize step adds these new insights into a workflow. This can include replacing labor-intensive and often mundane tasks like manually compiling mass volumes of data with automation, contextual routing, summarized analysis, and creating intelligence dashboards. 

The last step is what keeps the process going, which seems counter-intuitive, but the executive learns it is a feature of NLP. Once new insights are put in place to realize achievable outcomes, this new data is used to expand on and repeat for further analysis that will result in a deeper understanding.

While these concepts require a basic understanding of NLP, the eight steps succinctly sum up the process. The executive has developed a better understanding of how NLP can positively impact their bottom line. 

Primer strives to help the world understand the power of NLP and what it can do to help businesses make better decisions and gain a competitive advantage. The “8 Steps to Get Started with NLP” is one of myriad efforts to pique interest, start conversations, and educate the business community.

For more information about Primer and to access product demos, contact Primer here.

Imagine logging into your work computer one Wednesday morning and seeing untrue social media posts claiming the company you work for is a fraud.  Simultaneously you and your colleagues receive a report from an unknown source presented as comprehensive research warning investors against the company because it is a fraud. Several weeks pass before it is discovered that the company who published the report is a shell with no discernible employees and operating from an unknown address across the world.  But why would they write this report? Was the entire company created just to spread false information about your employer? 


Unfortunately the story above is not made-up. It’s also becoming less of an anomaly, especially in the crypto industry. Spreading disinformation in the crypto industry is prevalent and persistent and it often intermingles with real investment concerns.  The promulgation of disinformation with fear, uncertainty, and doubt or FUD,  is intended to confuse investors and potential investors.  Questions around hot button issues can be made intentionally to illicit FUD in an effort to affect the associated token’s price and popularity. The concept of FUD has become so pervasive that crypto sector social media users will use “FUD” as a word to call attention to any posts that negatively portray a crypto project.

AI Machine Learning tools can help to detect disinformation campaigns

New advancements in AI/Machine Learning, specifically Natural Language Processing (NLP), can help detect disinformation and synthetic text as well as partition the claim and the counterclaim of a disinformation campaign. This allows crypto projects to quickly see what is being said on each side of a dispute. 

With Command crypto companies can see the disputed information flagged for each report in the feeds column. They can also get perspective on the amount of FUD they are facing compared to others in the space. Additionally, Command displays FUD trends overtime and categorizes the organizations and people discussed in the posts. This helps in conducting investigations into the targets of the post and who is behind the disinformation campaign.

How pervasive is FUD?

FUD around crypto projects tends to focus on what governments will do about it. This has largely stemmed from China’s decision to ban crypto transactions and mining. This FUD gets recirculated frequently as China reaffirms its decision or cracks down on underground mining noting concerns about energy use. Creating a recent spike in FUD claims is the intensifying scrutiny of blockchain assets by the Securities and Exchange Commission and other U.S. regulators


Disinformation peddlers, in the form of bots or paid influencers, tend to pile on top of these fears with statements like those in the image below.  This social media influencer is known by many in the crypto sector to consistently post negative information about Tether and Bitcoin. He used the press release to support his campaign against both companies. Notably, the statements referenced in the post never mentioned Bitcoin or Tether. They focused on the impact mass adoption of stablecoins would have on traditional financial markets.

Disinformation in the crypto sector tends to skyrocket with any downturn in the token price. Take Ethereum (token: ETH) as an example. The first chart below shows ETH price in December 2021. The second chart shows a spike in FUD statements at the end of December when the price of ETH had its most severe decline.

In looking at the results from a basic Twitter search for the terms “FUD” and any of the top 20 crypto companies over the month of December there are 254 hits. Likewise, for Reddit there were 71 hits. While these numbers might not be alarming it’s important to note that they are only scratching the surface. This is because when social media users post FUD they don’t usually flag the term. This search is most often capturing other users pointing out FUD in other posts. This search also doesn’t cover discussions in threads of posts.

FUD contributes to market volatility, brand bullying

One of the oft-cited reasons for not investing in crypto companies is because of volatility. In November 2021, for example, Beijing reiterated its stance against Bitcoin miners which likely contributed to a crypto selloff over the next several days. The price of Bitcoin dipped 2.9% and Ethereum and Solana dropped 4.6% and 6.7%, respectively, following the statements.

The crypto industry is largely unregulated. And the federal government, for the most part, appears to still be figuring out how it all works. Couple the lack of oversight with the fact that most people interested in this sector shy away from central authorities. As a result many of the victims of FUD do not see legal recourse as an option.

Instead of court battles, they have taken to relying on community advocates to counter the messaging. These are paid and unpaid influencers who are supposed to support the brand and raise awareness about new developments through social media and educational meet-ups. Ripple has its XRPArmy, Chainlink has LINKMarines, and Dogecoin has the DOGEArmy, just to name a few. 

Yet more often these advocates are needed to focus on identifying and squashing false information directed at the brand. Because these are people financially invested in the company they take it too far and can contribute to brand degradation by attacking anyone questioning the project. Thus putting them directly at odds with their original purpose. 

The XRP Army, for example, is known for its scale and organization. If someone posts FUD about Ripple/XRP a foot soldier will spot the tweet and rally the troops by tagging the #XRPArmy. Next a flood of accounts will “brigade” the alleged FUD-monger, posting dozens or hundreds of comments. The attack comes in the form of an inundation of thousands and thousands of angry notifications that lasts for days.

Originators of FUD campaigns are difficult to identify

FUD campaigns are often hard to trace back to the originator because they will use fake companies and bots to amplify their message. And the cost of using bots to synthetically amplify content is relatively cheap. The New York Times in 2018 found that 1,000 high-quality, English-language bots with photos costs a little more than a dollar. See the possible bots intermixed with human posts below intensifying questions about whether it is time to sell Cardano’s ADA token below.

New synthetic text capabilities will make FUD campaigns even harder to trace

Bots are often detectable because they are posting the same message over and over. When you look at a bots profile they often have ‘tells’ such as imperfect use of the language, appear to have a singular theme to their posts, and have numerous bot followers. 

But these ‘tells’ are going to get increasingly difficult to identify with recent advancements in synthetic text generation. Last March researchers in the U.S. open sourced GPT-NEO, for the first time, making available to the public a next-generation language model. With the advent of these new generation language models launching a FUD campaign to try to drag down a competitor’s brand or for a short campaign will be even more difficult to detect. In fact, last summer, ​a team of disinformation experts demonstrated how effectively these algorithms could be used to mislead and misinform. The results are detailed in this WIRED article and suggest that it could amplify some forms of deception that would be especially difficult to spot.

Primer’s NLP Engines can help detect synthetic text and disinformation

Rather than continuing to invest in defensive armies or influencers to detect and flag FUD peddlers, the crypto space could benefit from an automated solution leveraging AI. Primer Command does all of this. Command ingests news and social media feeds and automatically detects, flags, and displays the origin and source of disputed information. This enables users to understand its provenance and evaluate its accuracy. This additional context also provides early warning and a means to constantly monitor the information landscape.

Command can also classify text as likely to have been written by a machine. It does this by automatically evaluating 20 different user signals to flag automated or inauthentic social media accounts.  Signals include how quickly an account is gaining followers, how many accounts they’re following, the age of the account, and even the composition of the account name. This information would allow crypto companies to evaluate the posts’ accuracy.

These tools hold more promise than manual efforts because they are impartial, within the parameters of how they are designed. It is an algorithm that identifies the FUD instead of someone with a stake in the project’s success. This is critical to assist in neutralizing the adversary without stoking the flames with numerous negative posts. By automating the identification of FUD campaigns, the project’s community can get back to focusing on brand promotion and education.  

Learn More

For a free trial of Primer Command or to learn more about Primer’s technology, Ask for Demo or Contact sales  to discuss your specific needs. You can also stay connected on Linkedin and Twitter.

“We create the tools behind the decisions that change the world. ©2022 Primer”

The NLP revolution kicked off in 2018 when BERT dropped. In 2019 and 2020, it was one mind-bending breakthrough after another. Generative language models began writing long-form text on demand with human-level fluency. Our benchmarks of machine intelligence — SNLI, SQuAD, GLUE — got clobbered. Many wondered if we were measuring the wrong things with the wrong data. (Spoiler: yes.) Suddenly, we no longer needed tens of thousands of training examples to get decent performance on an NLP task. We could pull it off with mere hundreds. And for some NLP tasks, no training data was needed at all. Zero-shot was suddenly a thing.

Over the past year, NLP has cooled down. Transformer architecture, self-supervised pre-training on internet text — it seems like this paradigm is here to stay. Welcome to the normal science phase of NLP. So, what now?

Time to industrialize

Efficiency, efficiency, efficiency, and more efficiency. The cost of training and serving giant language models is outstripping Moore’s law. So far, the bigger the model, the better the performance. But if you can’t fine-tune these models on fresh data for new NLP tasks, then they are only as good as your prompt engineering. (And trust me, prompt “engineering” is more art than science.) The focus on efficiency is all about industrializing NLP.

“The revolution, in the style of Kazimir Malevich” by John Bohannon using CLIP+VQGAN

The NLP benchmarks crisis is another sign that we’re in the normal science phase. Here’s an open dirty secret: The purpose of today’s NLP benchmarks is to publish papers, not to make progress toward practical NLP. We are getting newer, better benchmarks — Big-Bench, MMMLU — but we likely need an entirely different approach. The Dynabench effort is an interesting twist. But until we stop using performance metrics known to be broken — looking at you, ROUGE — industrial NLP will be harder to achieve.

Normal science is applied science. We are moving from “Wow, look at the impressively fluent garbage text my model generates!” to “Wow, I just solved a million-dollar NLP problem for a customer in 2 sprints!”

NLP is eating ML

But meanwhile, another revolution is brewing on the edges of NLP. The phrase “Software is eating the world” became “ML is eating software.” Now it looks like “NLP is eating ML,” starting with computer vision.

The first nibble was DALL-E. Remember when you first saw those AI-generated avocado armchairs? That was cool. But then along came CLIP+VQGAN. And just last month, GLIDE made it possible to iteratively edit images generated by text prompts with yet more text prompts. This could be the future of Photoshop.

The trend is the merging of cognitive modalities into a single latent space. Today, it’s language and photos. Soon, it will be language and video. Why not joint representation of language, vision, and bodily movement? Imagine generating a dancing robot from a text prompt. That’s coming.

Go beyond English

Bilingual language models lost the throne in 2021. For the first time, a multilingual model beat them all at WMT. That was a big moment that surprised no one. Machine translation has caught up with the multi-task trend seen across the rest of NLP. Learning one task often helps a model do others, because no task in NLP is truly orthogonal to the rest.

The bigger deal is the explosion in non-English NLP. Chinese is on top — no surprise given the massive investment. There are now multiple Chinese language models as big or bigger than GPT-3, including Yuan and WuDao. But we also have a Korean language model, along with a Korean version of GLUE (benchmark problems notwithstanding). And most exciting of all is an entirely new approach to multilingual NLP called XLS-R trained on a massive collection of speech recordings in 128 languages. This could become the foundation NLP model for low-resource languages.

Build responsibly

If 2020 was the year that NLP ethics blew up, then 2021 was the year that ethics considerations became not just normal but expected. NeurIPS led the way with ethics guidelines and a call on authors to include a “broader impact” section in their papers. 

But like the rest of the AI industry, NLP is still self-regulating. That will change. The most significant action in AI self-regulation last year was Facebook’s voluntary deletion of its facial recognition data. That move was surely made in anticipation of government regulation and liability. (They also have more than enough surveillance data on their customers to power their targeted advertising.)

Where will the rubber meet the road for NLP regulation? AI companies that work closely with governments will lead the way, by necessity. Here at Primer, we build NLP systems for the US, UK, and Australian governments, helping with tasks as diverse as briefing report generation and disinformation detection. This front-row seat to the NLP revolution is humbling.
One thing is now clear. The machine learning tools that we are building today to process our language will transform our society. An intense research effort is underway to spot potential biases, harms, and mitigations. But every voice of hope and concern for this technology must be heard. Building NLP responsibly is our collective duty.

Read more: The Business Implications of Machines that Read and Write

I have a strange new pandemic ritual. Every Monday, I gather my team of machine learning engineers in a video meeting to kick off the week. But before we get down to business, I use a giant neural network to generate questions that help us go deeper than small talk. We vote on our favorite question and then take turns answering it.

  • What would you do if you knew you could not fail?
  • If you were guaranteed that you would live forever, would you still be afraid of death?
  • What is the number one thing you wish you could change about how people see you?

Those are actual questions the model generated. They indeed lead to deep talk.

We’ve been working remotely at Primer since the earliest days of the pandemic. This ritual is one of our solutions to the problem of feeling disconnected to each other. I can tell you: It really works. 

Here’s how I do it. 

First, I create a prompt consisting of numbered questions. For example:

  1. What is the meaning of life?
  2. If you were the last person on Earth, who would you choose to join you?
  3. What animal would you be for a day?

You can find lists of such questions all over the internet, probably inspired originally by this 2015 New York Times article, 36 questions for a private conversation starter. I find that a list of at least 20 questions works best.

Then I feed this prompt to a big autoregressive language model. I use GPT-J-6B, the open source model created by Eleuther. I just copy-paste the prompt and run the model in Primer’s developer interface. The model’s inference takes about 5 seconds.

Photo credit: John Bohannon, generated with VQGAN+CLIP with the prompt “A beautiful rainbow in the style of van Gogh.”

What comes back is machine-written text that continues where mine left off:

4. What thing would you never let anyone see you do?
5. What color are your dreams?
6. If you could have one superpower, what would it be?

Yes, these neural networks know how to count. My team uses the numbers to vote for their favorite.

I admit it may seem bizarre to use a machine to help people move beyond small talk and connect. But this is artificial intelligence at its best: Working together with humans to help them do what they do best. 

I generated 365 Deep Talk questions so you can try it yourself.

How mapping bottom-up climate action can drive more strategic climate solution responses and help us adapt thoughtfully. 

At the recent COP26 climate summit, a Minister from the Pacific island country of Tuvalu announced that it would be seeking to retain legal status as a country even if its entire territory were to become submerged [Reuters]. He was standing thigh-deep in the ocean in an area that used to be dry land. His speech made it clear—the effects of climate change are here today.

When it comes to tackling the climate crisis, we typically think about solar power, electric vehicles, and carbon capture to mitigate future climate change. But Tuvalu’s story is the tip of the iceberg of climate adaptation — the messier, less-defined problem of how human civilization will respond to the changes that are here now and predicted to accelerate over the next 30 yrs — even if mitigation efforts are successful.

In order to help define this messy space, Primer recently partnered with Vibrant Data Labs, a social impact data science group, to make sense of this broader and more diverse climate landscape. Crunchbase and Candid provided data on over 12,000 companies and nonprofits funded in the past 5 years that are addressing climate-related topics. Primer’s natural language processing (NLP) engines mined these organizations’ descriptions to generate one of the first-ever, data-driven conceptual hierarchy of topics to better understand the shape of our current response, and its potential gaps. This unique perspective comes bottom-up from how the private and social sector organizations on the ground describe what they do — not by what is most spoken about in the news or social media.

Our analysis suggests that while new technologies are emerging to address climate mitigation, existing organizations that have historically tackled structural inequities (e.g, gender equity, migrant rights, homelessness) are uniquely poised to address climate adaptation challenges which permeate every aspect of civil society. Our sample showed these organizations are beginning to add a climate lens to their work on diverse social issues.

Defining the Climate Space

We created a hierarchy of interrelated topics based on the company descriptions. Using this hierarchy, we are able to surface the broad topics in climate work and also drill down into specifics. 

Examining the topics in this way revealed there are two major branches: one dealing with topics related to preventative technologies (Environment And Renewable Energy) and the other with topics addressing the human impact of change (Public And Social Benefits).  This computational technique led to a close split between mitigation and adaptation. It’s exciting that our method could organize these topics in a way that gets close to how a human would do the task.

The topics underneath our mitigation branch (Environment and Renewable Energy), are what one may expect: “Water”, “Nature and Conservation”, and “Energy, Storage, and Energy Waste”. Adaptation work is much more varied and therefore harder to define. Our analysis can help us paint a crisper image of this emerging landscape.

This topic hierarchy shows organizational distribution on climate change topics, with higher convergence at the top vs the bottom.

Climate Change as a Social Issue

The Intergovernmental Panel on Climate Change (IPCC) defines adaptation as “the process of adjustment to actual or expected climate and its effects” [IPCC]. Using this definition, we see some top level branches that are climate adaptation. As the earth warms and extreme weather becomes the new normal, Disaster Relief and Preparedness will be critical to serving areas affected.

A less obvious topic might be Community And Neighborhood Development. The subtopics within it seem like quite standard areas pertaining to social issues such as: Health, Affordable Housing, Human Rights, Government Advocacy, and Gender Equality. Looking deeper into the language of these organizations we can see how they are incorporating a climate lens to their work.  For example, here is the description of one of the organizations working in Gender Equality:

MADRE is an international women’s human rights organization founded 37 years ago.  We partner with grassroots women to create social change in contexts of war, disaster and injustice. To advance women’s human rights, MADRE supports grassroots women to meet their basic needs as a bridge to develop their advocacy and political participation….

…. Our actions to confront climate change today will decide the futures of our planet and of generations to come. You can join the women leading the way.    Climate change is a global threat, and poor, rural and Indigenous women are hardest hit. They are impacted first and worst by the food shortages, droughts, floods and diseases linked to this growing danger. But they are more than victims. They are sources of solutions, inventing innovative, locally-rooted responses.    Through our Women Climate Defenders initiative, MADRE’s grassroots partners enable their communities to adapt to climate change. They build clean water systems to guard against drought and seed banks to preserve future harvests.

This is an example of how an organization that has been addressing women’s human rights for 37 years can contribute today to building climate resilience in the most vulnerable communities. It also highlights how climate adaptation requires addressing diverse, interdependent topics.

We can dive deeper into the Gender Equality data to understand the key topics that organizations in this field are working on today. A quick glance at this chart shows a wide range and diversity of topics in the climate adaptation cohort, from Human Rights to Infrastructure to Youth Organizing and Antidiscrimination.

The topics which co-occur most frequently with gender equality cover a range of socially minded topics which are not all tightly related to gender equality.

Let’s compare it to a topic from our climate mitigation set, Nature Conservation and Environment

The topics which co-occur most frequently with nature conservation and environment are very conceptually similar and mostly related to climate mitigation.

Organizations in this cohort work on Sustainability, Renewables, Water Conservation, Sustainable Agriculture, and Wildlife Preservation. It seems that most of these issues are more proximate to each other.

To further peel back the layers on current climate solutions, let’s take a deeper look at “crowding” or “spreading” of focus areas by organization. With NLP, we can approximately measure organizational “topic coherence” which tells us if a given organization optimizes on breadth or depth, and exactly how far apart the topics are within that cohort. We created a score from 0 to 1 that calculates how similar an organization’s topics are to each other—we call this the “organization focus score”. Organizations that focus on a narrow set of topics will have scores closer to 0. We can then extrapolate to the topic level to measure how narrowly focused the organizations in each topic are. When we plot this out from 0 to 1, we see topics relating to climate adaptation (Public and Social Benefit) are being addressed by organizations that are more broadly focused than the organizations addressing climate mitigation (Environment and Renewable Energy) topics.

The topic coherence score measures how closely related the topics a given organization works in are to each other. A score of 0 closer to 0 indicates the topics are very similar and a score closer to 1 means they are all very dissimilar. Climate adaptation topics (Public and Social Benefit) contain organizations with a more diverse set of focus areas than the climate mitigation topics.

Our analysis reveals that, while an organization working on mitigation will typically be working on a single, defined solution, organizations working on climate adaptation are fighting on multiple fronts.  


“In an interconnected world, it is exactly this messiness that funders need to embrace”, says Vibrant Data Labs’s Eric Berlow. “Traditional venture capital tends to fund focused, scalable solutions, with easy-to-measure outcomes, like renewable energy. But the climate crisis is an ‘all hands on deck problem’. Winning on one corner of the problem is an important piece; but if structural and systemic inequities in climate adaptation are not addressed, like the people of Tuvalu above, we all lose. We all feel the climate impacts of supply chain shocks, forced migration, civil unrest, and war. The most recent IPCC report suggests these trends will exacerbate over the next 30 years even if renewables and carbon capture solutions are successful. Climate funders will have to adopt a more holistic multi-pronged approach to rise to this challenge.”

Conclusion

As climate change becomes more and more a central part of our lives, understanding the landscape of solutions and providers gives us perspective on the magnitude of the space. We used NLP to analyze the work of over 12,000 companies to better understand where private and public organizations were focusing their efforts. In doing so, we highlighted the broad set of topics that are climate related and illustrated that many organizations working across the diverse social sector are now adding climate solutions to their efforts to enhance equity and resilience.

In a coming post, we will present our partner, Vibrant Data Labs’ story in which they take this analysis a step further to highlight the solution areas that are receiving the most funding.

Natural language processing (NLP) automates the work that would previously have required hundreds of researchers. These machines read and write – and they’re changing the future of business.

With NLP, we can now analyze information at machine speed, but with human-level precision. As reading and writing is automated through machine learning, companies are freed up to focus on the challenges that are uniquely human.

Read more here