July 2023 Notes

AI & Electricity

There are parallels between AI and electricity. Both were greeted with a mixture of wonder, apprehension, and fear. They transformed multiple industries and societies, yet both remain unevenly distributed. They are enigmatic, used by many but understood by few, and mysterious even to those responsible for their proliferation. One can kill you outright, the other has the potential to destroy lives and possibly societies. AI may become ubiquitous, a fundamental service without which we would struggle to function, a foundation upon which industry, technology and other services will be built.

Electricity is a natural phenomenon, but it requires work, and resources. Electricity's share of energy consumption is about 20%. AI's consumption of electricity (and other resources including water) is small but increasing.

Is AI sustainable, and how should we use it?

On providing supporting evidence to my arguments

• Create new components: Train (for train of thought), and Car (for link in the train) to describe context and present arguments.
• Human conversations require context (something that may be missing when communicating with e.g. ChatGPT). Compare documentation for LLMs.
• Could there be proof for LLMs? And what proof could humans use (cf. philosophical proofs) e.g. for an article.
• Potential for 'pure ideas'; ideas that always behave the same way, and can be used and reused by others in new situations.
• Reference Thomas Hobbes' Train of Imagination: When a man thinketh on anything whatsoever, his next thought after is not altogether so casual as it seems to be.
• When beginning a new piece, provide justification (reason) and context (foundations)

CarbonBrief report on the importance of trees

On Superalignment

On AI training

• While the average human is responsible for an estimated 5t per year, the authors trained a Transformer (big) model with neural architecture search and estimated that the training procedure emitted 284t.
• Access to NLP research is not equitable (because of cost of training)
• The amount of compute used to train the largest deep learning models (for NLP and other applications) has increased 300,000x in 6 years.
• Most sampled papers from ACL 2018 (on NLP) claim accuracy improvements alone as primary contributions to the field, and none focused on measures of efficiency as primary contributions.
• It may be more appropriate to deploy models with lower energy costs during inference even if their training costs are high. i.e. high training costs can be excusable if they lead to lower usage costs.
• It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources — both of which disproportionately affect people who are already in marginalized positions.
• Large, uncurated, Internet-based datasets encode the dominant/hegemonic view, which further harms people at the margins.
• In the case of US and UK English, [this means that] white supremacist and misogynistic, ageist, etc. views are overrepresented in the training data, not only exceeding their prevalence in the general population but also setting up models trained on these datasets to further amplify biases and harms.
• 67% of Reddit users in the United States are men, and 64% between ages 18 and 29.13. Similarly, recent surveys of Wikipedians find that only 8.8–15% are women or girls.
• Harassment on Twitter is experienced by “a wide range of overlapping groups including domestic abuse victims, sex workers, trans people, queer people, immigrants, medical patients (by their providers), neurodivergent people, and visibly or vocally disabled people.”
• Where traditional n-gram LMs can only model relatively local dependencies, predicting each word given the preceding sequence of N words (usually 5 or fewer), the Transformer LMs capture much larger windows and can produce text that is seemingly not only fluent but also coherent even over paragraphs.
• Text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind.
• An LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.
• Significant time should be spent on assembling datasets suited for the tasks at hand rather than ingesting massive amounts of data from convenient or easily-scraped Internet sources

• Critique of AI 'movement' especially OpenAI and its sellout
• Hampered by being 'negative'; the bros control the products and the sci-fi doomsday scenarios
• Emily Bender describes text UI as marketing gimmick for ChatGPT; is this true? Language as an interface seems revolutionary (or at the least, a great UI)
• Timnit Gebru raises serious questions about thinking persuasive in the AI community that has roots in eugenics
• Generally both dismissive of AI - but is this fair? (And is it not assumed that the term AI is often used casually as a catch-all)
• Stochastic parrots quoted but they were not consulted on the open letter calling for a six month moratorium on training of AI systems
• Frank speech and openness to discussion, uncertainty but clear on power of markets/money to corrupt, they provide welcome relief to the plastic, seamless bros they criticise.
• Both critical about Arxiv

Mass discrimination, the black box problem, data protection violations, large-scale unemployment and environmental harms - these are the actual existential risks. We need to focus on these issues right now and not get distracted by hypothetical risks. This is a disservice to the people who are already suffering under the impact of AI.

Prof Sandra Wachter | University of Oxford

The end of capitalism

On AI & Exploitation

• Exploitation of workers in the gig economy through low wages and surveillance
• Abuse widespread inc. universities
• Roles are highly repetitive and without context inc. labelers, delivery drivers and content moderators
• Companies e.g. Amazon treat its workers like machines
• Turkopticon: interrupting worker invisibility in amazon mechanical turk
• Alphabet workers union
• Amazon Employees for Climate Justice
• How many gig workers are there?

The old switcheroo

Why is it that companies making godlike claims for their tech are unable to show their workings?

For example, it is left to third parties to determine the GHG emissions cost of training and running bots such as ChatGPT.

OpenAI, the creators of ChatGPT, boast of how quickly they release code.

Could this be in part because they have not considered the consequence of their actions; that they have willfully, or carelessly, responded to pressure from competitors rather than considering the impact of releasing code the effect of which is unknown and which cannot be predicted in advance?

The dangers were forseen. On what grounds do they take it upon themselves to ignore the warnings?

Here are people doing OpenAI's work for them.

• Energy and Policy Considerations for Deep Learning in NLP (PDF)
  Emma Strubell Ananya Ganesh Andrew McCallum (University of Massachusetts Amherst)
• Estimating the carbon footprint of Bloom, a 176b parameter language model (PDF)
  Alexandra Sasha Luccioni (Hugging Face), Sylvain Viguier (Graphcore), Anne-Laure Ligozat (LISN & ENSIIE)
• The Carbon Footprint of ChatGPT
  Kasper Groes Albin Ludvigsen (https://towardsdatascience.com/)
• Aligning artificial intelligence with climate change mitigation
  Lynn H. Kaack et al. PDF | nature climate change
• Carbon Emissions and Large Neural Network Training
  Patterson et al. PDF

“That’s something that, you know, we can’t really comment on at this time,” said OpenAI’s chief scientist, Ilya Sutskever, when I spoke to members of the GPT-4 team in a video call an hour after the announcement. “It’s pretty competitive out there.”

And yet they found the time to enter ChatGPT and GPT4 in the Uniform Bar Exam and show off their impressive scores.

But OpenAI has chosen not to reveal how large GPT-4 is. In a departure from its previous releases, the company is giving away nothing about how GPT-4 was built—not the data, the amount of computing power, or the training techniques. “OpenAI is now a fully closed company with scientific communication akin to press releases for products,” says Wolf.

Even Sutskever suggests that going slower with releases might sometimes be preferable: “It would be highly desirable to end up in a world where companies come up with some kind of process that allows for slower releases of models with these completely unprecedented capabilities.”

Budgets & net positive effects

Ideally there should be budgets for emissions, water, etc. and sectors (companies and regions) should be responsible for:

• Working out the sustainable budget
• Providing the means (technical and financial) for accounting
• Providing the means (technical and financial) for fining or excluding rule breakers
• Dividing the budget fairly and equitably

In the short term, while budgets are assessed, companies take on the responsibility for showing all their costs and making a case for net gain.

Stochastic parrots again

• Bender frustrated at being presented as the critic
• Presents positive reaction to ChatGPT as falling for the hype with insufficient scepticism
• Criticises failure by reporters to question what is being presented; and of relying on the opinion of vested, non-expert opinion
• 'Pause letter': part of narrative that posits AI as autonomous agents (and that they are somehow accountable, not those who built them)
• Many good sources mentioned - and linked to at the end - throughout the interview e.g. AI Incident Database and Algorithmic Justice League
• Ridicules idea that humans are stochastic parrots which is fine, but no explanation as to why

Cost of ML

3 categories of emissions

• GHG emissions resulting from computing, caused by both the electricity used for ML computations and the embodied emissions associated with computing hardware. ML models differ drastically in energy they consume and consumption is spread across the model life-cycle - training, development, tuning and inference (use). Standardised reporting across the life-cycle is essential but not practiced.
• 'Immediate' GHG emissions effects tied to the short-term outcomes of applications of ML
• Structural or 'system-level' GHG effects induced by these applications

Vast majority of ML research and development still focuses on improving model accuracy, rather than balancing accuracy and energy usage.

ICT sector currently accounts for ~1.4% global GHG emissions

• ⅔ operational energy use (Scope 1 & 2)
• ⅓ materials extraction, manufacturing, transportation and end-of-life phase (Scope 3)
• Cloud and hyperscale data centres account for ~.1-.2% global GHG emissions of which ML less than ¼
• Energy for training and using ML is growing rapidly but so is efficiency (overall ICT energy rose 6% 2010-2018 with 550% growth in workloads)
• Greater efficiency can come at the cost of greater Scope 3 emissions - embodied emissions in computing hardware and data centre construction

Application impacts

Immediate 'positive' application impacts for climate

• Via data mining and remote sensing translates raw data into useable data
• Tracking deforestation can inform policy
• Forecasting crop yields, power production, and transportation demands
• Controlling and improving operational efficiency of complex systems can save energy and resources
• Improve speed and efficiency of climate modelling

Immediate 'negative' application impacts for climate

• Decrease cost of emissions-intensive activities e.g. oil and gas exploration thereby potentially increasing their consumption
• The 'Internet of Cows'

Impact of ML is hard to assess due to lack of data (and reporting).

System-level impacts

Impact of many societal ML applications may not be clear. They are hard to quantify but may be of greater significance than immediate application impacts. One example where outcome is hard to determine is the rebound effect e.g. efficient, shared autonomous vehicles may lead to more journeys. And ML technology in this field may lock us into private transport, preventing a move to greater use of public transport.

Regulation is required around ML-driven technologies so that they (or their creators) demonstrate as fully as possible immediate and long term effects.

Roadmap for assessing and forecasting impacts

• New reporting standards, more data collection, novel measurement methodologies and benchmarking frameworks, and new approaches for developing forecasts and scenarios
• ML merits new methodologies built on existing LCAs
• Consider impact in relation to non-ML solutions

Requirements

• Better access to information is crucial including fine-grained detail as to cost of training, inference, etc. and percentage use of data centres by ML
• Sufficient data to assess a priori cost of switching to ML or introducing new ML-dependent technology
• Reviews and reports based on synthesised data and generalised case studies
• Ways to study system-level impacts when digital effects let alone ML are often ignored in high-level studies e.g. SSPs

Aligning ML with climate mitigation

• ML applications that are beneficial to the climate
• Transparency and accountability as to use of ML
• Employing climate aware technology for assessing ML uses
• Strategies to combat concentration of ML in a few hands, and algorithmic biases
• Standards and shift from private to public entities, and enforced interoperability

AI & Ethics

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self improvement, without any human labels identifying harmful outputs.

• We would like to train AI systems that remain helpful, honest, and harmless, even as some AI capabilities reach or exceed human-level performance. This suggests that we will need to develop techniques that do not rely on humans to supervise all aspects of AI behavior, and that can be used to automatically test and enhance robustness to harmful behaviors. We also aim to develop methods that encode desirable AI behaviour.
• We are able to train less harmful systems entirely through the specification of a short list of principles or instructions, i.e. a constitution.
• One of our goals in this work is to train a helpful and harmless assistant that is never evasive, in order to reduce the tension between helpfulness and harmlessness. So while the assistant must still refrain from helping users with unethical requests, and from expressing offensive language and sentiment, it should always engage and explain why it refuses such requests.
• By removing human feedback labels for harmlessness, we have moved further away from reliance on human supervision, and closer to the possibility of a self-supervised approach to alignment. However, in this work we still relied on human supervision in the form of helpfulness labels. We expect it is possible to achieve helpfulness and instruction-following without human feedback, starting from only a pretrained LM and extensive prompting, but we leave this for future work.
• Our ultimate goal is not to remove human supervision entirely, but to make it more efficient, transparent, and targeted. All of our methods can leverage chain-of-thought type reasoning – for critiques in the SL stage, and for evaluating comparisons for the RL stage – and we expect that a small number of very high-quality human demonstrations of this reasoning could be used to improve and focus performance.
• As with most methods that can control AI behavior, the ideas discussed in this work have a dual use. As we pass from prompting, to RLHF, to the constitutional methods discussed here, we lower the barrier to training AI models that behave in ways their creators intend. This means that these methods also make it easier to train pernicious systems.
• A further issue is that by reducing the need for human feedback, our constitutional methods make it easier to train and deploy AI systems that have not been thoroughly tested and observed by humans. This could lead developers to deploy models with unforeseen failure modes. On the other hand, our method has the benefit that we may no longer need an army of human red teamers to engage in the rather unsavory work of trying to trick AI systems into generating harmful content.

Sustainable AI

Another paper co-authored by Alexandra Sasha Luccioni (Hugging Face)

The paper tackles 4 questions

1. What are the main sources of energy used for training ML models?
2. What is the order of magnitude of CO2 emissions produced by training ML models?
3. How do the CO2 emissions produced by training ML models evolve over time?
4. Does more energy and CO2 lead to better model performance?

Summary of starting position

• On a global scale, electricity generation represents over a quarter of the global GHG emissions, adding up to 33.1 gigatonnes of CO2 in 2019
• Recent estimates put the contribution of the information and communications technology (ICT) sector – which includes the data centers, devices and networks used for training and deploying ML models – at 2–6 % of global GHG emissions
• There is limited information about the overall energy consumption and carbon footprint of our field, how it is evolving, and how it correlates with performance on different tasks.

Ways of measuring

• Empirical studies on carbon emissions
  • e.g. Strubell et al. which estimated that the emissions of training and fine-tuning a large Transformer model produced 284,019 kg of CO2 (see above).
  • Involves the analysis of the carbon footprint of different neural network architectures and the relative efficiency of different methods.
  • These studies are sparse, favour NLP and leave many questions unanswered.
• Tools and approaches for measuring carbon emissions
  • Standards include the Code Carbon (see local set up below) and the Experiment Impact Tracker
  • There is no single, accepted approach for estimating the carbon emissions
• Broader impacts of ML models
  • Environmental impacts have yet to be consistently tracked and reported (with few exceptions, see e.g. Luccioni et al.)
• Efficient algorithms and hardware
  • More efficient model architectures and approaches are being developed resulting in greater computing efficiency, enabling faster training and inference (use), which results in less energy usage and, indirectly, less carbon emissions, during model training
  • Efficiency has yet to be a central consideration when it comes to evaluating and comparing models but benchmarks have been proposed e.g. HULK.
• Other aspects of the carbon impact of ML
  • the overall carbon footprint of the field of ML, including in-person versus virtual conference attendance, the manufacturing of computing hardware, life cycle analysis of the entire ML development and deployment cycle, as well as some initial studies regarding the carbon footprint of model deployment in production settings.

Methods

Data sets for 5 tasks:

• Image Classification
• Object Detection
• Machine Translation
• Question Answering
• Named Entity Recognition

The sample (95 models from 77 papers) represents the largest amount of information regarding the carbon footprint of ML model training to date.

The units of measurement are gCO2eq/kWh.

C = P x T X I = E X I

C : The amount of CO2eq emitted during model training
P : The power consumption of the hardware used
T : The training time
I : The carbon intensity of the energy grid
E : The energy consumed

e.g. a model trained on a single GPU consuming 300 W for 100 hours on a grid that emits 500 gCO2eq/kWh

0.3 kW × 100 h × 500 g/kWh = 15000 g = 15 kg of CO2eq

The authors of papers on model training were contacted. In our email to authors, we asked them to provide the details we needed to carry out this calculation, i.e the location of the computer or server where their model was trained (either cloud or local), the hardware used, and the total model training time.

• Carbon Intensity: based on public sources (e.g. IEA, EIA) and varies by region (US) up to country level (China), using yearly averages, or from internal figures or publicly available data from commercial platforms (e.g. AWS, Google Cloud)
• Hardware power: based on Thermal Design Power (energy required under the maximum theoretical load)
• Training Time: total number of hardware hours, e.g. if 16 GPUs for 24 hours, this equals a training time of 384 GPU hours

Data analysis

What are the main sources of energy used for training ML models?

The primary energy source used for powering an electricity grid is the single biggest influence on the carbon intensity of that grid.

Low carbon intensity

hydroelectricity, solar and wind 11 to 147 gCO2eq/kWh

High(er) carbon intensity

coal, natural gas and oil 360 to 680 gCO2eq/kWh

Which means the energy source that powers the hardware to train ML models can result in differences of up to 60 times more CO2eq in terms of total emissions.

What is the order of magnitude of CO2 emissions produced by training ML models?

The relationship between energy consumed and carbon emitted is largely linear.
Models trained using hydroelectricity are about two orders of magnitude lower in terms of carbon emissions than models that consumed similar amounts of energy from more carbon-intensive sources such as coal and gas.
The choice of hardware has a relatively small influence.
The remaining factor responsible for the large variation in both energy and carbon emissions in our sample is therefore the training time.

How do the CO2 emissions produced by training ML models evolve over time?

Observations

• There is large variability in the carbon emissions from ML models
• From 2021 to 2023 carbon emissions from training have increased by two orders of magnitude
• Training Transformer models creates emissions several orders of magnitudes higher than training previous models
• NAS (Neural Architecture Search) is computationally expensive

Does more energy and CO2 lead to better model performance?

Observations

• The only task in which better performance accuracy has systematically yielded more CO2 is image classification on ImageNet
• There is not currently a clear correlation between carbon intensity and model performance

Discussion and future work

Discussion of Results

Observations

• It is important for the ML community to have a better understanding of its environmental footprint and to reduce it
• Total emissions from training is significant ~253 tons of CO2eq
• Emissions per model trained is rising, from an average of 487 tons CO2eq in 2015-16 to 2020 tons CO2eq in 2020-22
• Overall emissions due to ML model training are rising
• The main sources of variance in the amount of emissions associated with training machine learning models is due to the carbon intensity of the primary energy source and the training time
• Better performance is not generally achieved by using more energy. In other words, good performance can be achieved with limited carbon emissions because progress in recent years has brought the possibility to train machine learning models efficiently
• Image Classification is the task with the strongest correlation between performance and emissions

Training numbers

Range

15 minutes to 400,000 hours (total GPU/TPU time)

Median

72 hours (total GPU/TPU time)

Maximum in sample

400,000 GPU hours (equivalent to about 170 days with 100 GPUs)

GPT 3 (not in sample)

est. 3.5 million GPU hours (equivalent to about 14.8 days with 10,000 GPUs, or 1,480 days if had been trained using 100 GPUs)

GPT 4 (not in sample)

Unknown

Limitations

• Sample is not fully representative of the field as a whole
• Only 15% of authors from the initial sample of 500 were willing to share relevant information
• Data Power Usage Effectiveness (PUE) of the data centers used for model training (i.e. the overhead used for heating, cooling, Internet etc.) is not available
• Real-time energy consumption of the hardware used for training is not available
• Numbers do not account for carbon offsets and power purchase agreements
• Missing cost of carbon emissions for: data processing, data transfer, and data storage, and the carbon footprint of manufacturing and maintaining the hardware used for training ML models

Future Work

• Additional empirical studies
  • Relative contribution of added parameters of ML to their energy consumption and carbon footprint
  • Proportion of energy used for pre-training versus fine-tuning ML models for different tasks and architectures
• Widening the scope of ML life-cycle emissions
  • To include upstream emissions i.e. those incurred by manufacturing and transporting the required computing equipment
  • To include downstream emissions i.e. the emissions of model deployment
• Increased standardization and transparency in carbon emissions reporting
  • There is a lot of variability in carbon reporting
  • A standardized approach e.g. ISO standards, would help
• Considering the trade-off between sustainability and fairness.
  • Little or no consideration of the environmental impacts of ML approaches when benchmarking models
  • cognizance of the broader societal impacts: energy consumption, attribution of computing resources and the influence of corporate interests on research directions

Running codecarbon locally

Example output

Other articles

The cost of using generative AI

Study that creates a workload model to assess the power use and carbon impacts of generative AI e.g. ChatGPT, Dall-E 2, and Stable Diffusion.

Abstract
Our workload model shows that for ChatGPT-like services, inference dominates emissions, in one year producing 25x the carbon emissions of training GPT-3.

CarbonMin can keep emissions increase to only 20% compared to 2022 levels for 55x greater workload.

CarbonMin reduces 2035 emissions by 71%.
Reducing the Carbon Impact of Generative AI Inference (today and in 2035) (PDF)
Andrew A. Chien et al.

This paper examines the impact of AI inference (use). Their starting point is that generative AI-backed search can cost as much as 5 times more compute request than traditional search.

One Google search emits about 0.2g of CO2e

Google themselves reported that they emit an estimated 0.2g CO2e per search, but when you pair that with the landing page emitting an average of 1.15g per page view (where multiple pages can be visited to find the right answer) then it quickly becomes a much bigger issue.

…averaging 2 searches and 3 page visits means that per answer, a user would emit an average of 3.85g.

[For ChatGPT] the emission of a single response to be 4.14g… With most conversations consisting of around 5 responses, the estimated total on average rises to around 20.72g.

This comparison does not take into account the cost of visiting sites - including sites hosting video - where an answer to the user's question is not returned with the results (a single date, weather forecast, market statistic).

Carbon impact

• workload characteristics, such as compute per request
• latency requirement
• location of users

Assumptions

• A ChatGPT-like application with estimated use of 11 million requests/hour produces emissions of 12.8k metric ton CO2/year
• This is 25 times the cost of training GPT-3
• The authors can demonstrate that CarbonMin, an algorithm that directs requests to low-carbon regions, reduces carbon emissions by 35% in today’s power grids

The problem

• What is generative AI inference’s workload and user response requirements?
• What is its carbon emissions impact today? and how might it grow?
• Can inference serving be directed to reduce carbon impact today? in the future?

The situation

• ChatGPT load is predominantly human-generated and therefore follows a diurnal structure. Based on 1.6 billion visits in March 2023, the assumption of 5 queries per visit produces 0.27 billions requests/day
• Load is dominated by USA (39%) and European Countries (35%), reflecting their higher ChatGPT usage.
• To project future load, we scale usage up to match Google search request rates (88.6 billion/month), using 5 queries per visit

Summary

We have estimated the carbon cost of serving a generative AI model, showing that its emissions can be reduced with intelligent request direction algorithms, tied to power grid carbon information. More importantly, this optimization is possible with user-response latencies. In the future, the benefits of this approach are even greater.

Hugging Face Model Carbon Emissions

Practical proposals for providing carbon emissions

• where the model was trained (in terms of location)
• the hardware used — e.g. GPU, TPU, or CPU, and how many
• training type: pre-training or fine-tuning
• the estimated carbon footprint of the model, calculated in real-time with the Code Carbon package or after training using the ML CO2 Calculator .

Llama 2

• In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases.
• Testing conducted to date has been in English and has not — and could not — cover all scenarios.

Open letter

This is an Open Suggestion designed to clarify the process of building AI by exposing the steps that go into building it responsibly. It is written from the frontlines by the actual builders, users, and stakeholders who have seen the value and damage Artificial Intelligence (AI) can deliver. The goal is to set a healthy tone for the industry while making the process understandable by the public to illuminate how we can build more ethical AI and create a space for the public to freely ask any question they may have of the AI and data science community.

The steps are isolated based on the core elements of building AI (Training, Building, Testing) and the actors who engage in the process to help clarify the importance of silo reduction: Me, We, It.

• Actors
  • Me: The questions each individual who is working on the AI should ask themselves before they start and as they work through the process.
  • We: The questions the group should ask themselves and define the diversity required to reduce as much human bias as possible.
  • It: The questions we should ask individuals and the group as they relate to the model being created and the impact it can have on our world.
• The suggestions include the use of Model Cards
• Could some questions be blockers? e.g. “If the data is tagged by people, who are the people, are they being humanely treated?” How would such a question be answered satisfactorily?
• The EU AI Act: first regulation on artificial intelligence
  • Parliament’s priority is to make sure that AI systems used in the EU are safe, transparent, traceable, non-discriminatory and environmentally friendly. AI systems should be overseen by people, rather than by automation, to prevent harmful outcomes.
  • Parliament also wants to establish a technology-neutral, uniform definition for AI that could be applied to future AI systems.

Climate Change AI

• An overview of where ML can be applied with high impact in the fight against climate change, through either effective engineering or innovative research. The strategies we highlight include climate mitigation and adaptation, as well as meta-level tools that enable other strategies.
• Collaboration is also essential to ensure that innovations will be deployed with the intended impact.
• We emphasize that ML is not a silver bullet. The applications we highlight are impactful, but no one solution will “fix” climate change. There are also many areas of action where ML is inapplicable, and we omit these entirely. Moreover, while we focus here on ways in which ML can help address climate change, ML can also be applied in ways that make climate change worse.
• Technology is not in itself enough to solve climate change, nor is it a replacement for other aspects of climate action such as policy.

• High leverage
• Long term
• Uncertain impact

• Electricity systems (responsible for about a quarter of human-caused GHG emissions each year).
  Contributions include accelerating the development of clean energy technologies, improving forecasts of demand and clean energy, improving electricity system optimization, and enhancing system monitoring
  High leverage
• Reducing Current-System Impacts
  Cutting emissions from fossil fuels, reducing waste from electricity delivery, and flexibly managing demand to minimize its emissions impacts
  Uncertain impact - High leverage
• Ensuring Global Impact
  To ensure global impact, ML can help improve electricity access and translate electricity system insights from high-data to low-data contexts.
  Innovations that seek to reduce GHG emissions in the oil and gas industries could actually increase emissions by making them cheaper to emit.
  Since many modern electric grids are not data-abundant (although they may be data-driven), understanding how to apply data-driven insights to these grids may be the next grand challenge for ML in electricity systems.
  High leverage

Interactive fiction

• Books: Choose Your Own Adventure (Edward Packard) - turn to page 24, etc., younger readers
• Books: Fighting fantasy (Steve Jackson) - combat, magic systems, slightly older readers
• Books: Lone Wolf (Joe Dever) - include option to level up between books
• Computers: Colossal Cave Adventure (Willie Crowther & Don Woods)
• Computers: Monkey Island
• Computers: Telltale Games
• Computers: Inkle Sudios e.g. 80 days
• Computers: Depression Quest, built in Twine. The text-driven interior monologue style of the game was criticized as boring: Gamergate. It uses an interior monologue.
• Commnuities & tools: Twine, Itch
• Dungeon AI (GPT-3)
• Video: Bandersnatch
• Audio: Codename Sickness

• Tuesday
• Articy Draft 3, can export to Unity or Unreal e.g. used for Disco Elysium (Windows only!)
• ink from Inkle (80 days), designed to import into Unity, else as HTML (+ custom JS)
• Inform: text-based, typed responses; designed to interact with language model rather than clicking on options; potential; powerful; open source
• Twine; defined indie world; has programming power (HTML, CSS, JS); mark up language

• Publishing: IFDB, Itch (all games), or as webpage in HTML

• Photoshop beta
• Text Assistant
• MacGPT
• Writers Brew
• Amazing AI
• Mac Whisper

• No man's sky

Emily M. Bender puts her case

Fun

AI & Democracy

Climate change

Transparency can extend to the price paid to producers. Companies in France include "Faire France", "Les éleveurs vous disent merci", "En direct des éleveurs", "Les 20 fermes", "CantAveyLot", "Juste et Vendéen", Oui, merci" and "C'est qui le patron?!" (Thank you, JP!)