mrbusche.com

The Tipping Point of Change: How a Few Can Impact the Many

Forget the majority, it’s all about the minority.

In “Revenge of the Tipping Point,” Malcolm Gladwell explores the fascinating dynamics of social epidemics. He reveals how a small percentage of individuals can trigger widespread change, overturning the traditional notion that mass movements require mass participation.

The Rule of Thirds

Gladwell introduces the intriguing concept of “the magic of the third.” Through extensive research on group dynamics, he argues that a mere third of any group holds the power to drive significant shifts. This applies to everything from classroom integration to corporate boardrooms. A prime example? The transformative impact on workplace dynamics once women reached one-third representation in boardrooms.

The 1% Pollution Problem

The impact of a minority isn’t limited to social dynamics. Consider Denver’s air pollution problem. Nearly 40 years ago, Donald Stedman developed a device to measure vehicle emissions. While numerous studies highlight the potential of electric vehicles to reduce pollution, Denver uncovered a startling truth: if only the top 1% of polluting vehicles switched to electric power, the pollution reduction would be equivalent to having 10% of all cars go electric!

This isn’t to say we shouldn’t all strive to reduce our environmental footprint. However, it underscores the disproportionate impact of a small minority and the potential for rapid change if we can influence their behavior.

Super Spreaders: The COVID Case Study

The phenomenon of the influential minority extends even to the microscopic realm. A British study involving 36 volunteers, deliberately infected with COVID-19, revealed that a mere two individuals were responsible for a staggering 86% of the virus particles. This highlights the role of “super spreaders,” individuals who emit significantly more aerosol particles than average, in driving the pandemic.

The Takeaway: Focus on the Few to Impact the Many

Gladwell’s work offers a powerful lesson: change doesn’t always require convincing the masses. By identifying and influencing the key players - the “magic third,” the top polluters, the super spreaders - we can unlock a cascade of change and achieve our goals far more efficiently.

read more

Unhobbling an LLM refers to the process of removing limitations or constraints that prevent the LLM from reaching its full potential.

Imagine a really talented artist being forced to only use crayons. They could still draw some cool stuff, but imagine what they could do with paints and brushes or pencils.

The reasons for the hobbling are well intentioned. It’s kind of like keeping a powerful dog on a leash - you want to make sure things stay safe and under control.

Researchers are working on ways to “unhobble” LLMs so they can do every more amazing things. If you want to learn more about how they’re trying to make this happen check out this essay.

read more

Vector embeddings are a powerful tool for representing data in a way that machines can understand. But what if we could make them even better?

Binary quantized vector embeddings offer a way to do just that. By compressing vectors by 32x, we can achieve ~25x retrieval speedup while retaining 95+% accuracy. This means faster, more efficient search and analysis, all without sacrificing performance.

Check out this blog post to learn more about this exciting development: https://emschwartz.me/binary-vector-embeddings-are-so-cool/

read more

🎉 Exciting milestone: Our Legal Item Summarizer Gen AI application is now live! It’s been incredibly rewarding to work on this project due to the rapid feedback loop with our business partners and the immediate impact it’s having.

Context: Every day, new laws and regulations are enacted that can affect various areas of our business—from auto and pet insurance to financial compliance. Previously, our teams sifted through generic summaries to identify what mattered most for their business units, utilizing an RSS like feed for updates.

Our Solution: We leveraged Generative AI to deliver tailored, easy-to-read summaries that replace the one-size-fits-all approach. Now, business teams can quickly access the insights they need, focusing on higher-value tasks rather than combing through generic summaries.

For me, this is where Generative AI shines—transforming daily workflows and empowering people to focus on impactful work.

read more

The latest Thoughtworks Technology Radar is out. Here are a few blips that caught my eye:

• Function calling with LLMs - Integrate LLMs with external functions, APIs, or tools and allow them to act on their outputs
• Dynamic few-shot prompting - Dynamically include specific examples to guide the model
• Langfuse - Observability for LLMs
• Bruno - Open-source alternative to Postman
• ColPali - Leverages Vision Language Models to take a holistic approach for understanding text and visual content
• Instructor - Helping LLMs generate structured output effectively
• DeepEval - LLM evaluation framework but also offers some “red team” features

Check out the full report here: https://www.thoughtworks.com/radar

read more

Our school district recently hosted a book club discussion on “The Anxious Generation” by Jonathan Haidt, and it was an eye-opening experience for me and many other parents. The book dives deep into how smartphones and social media have significantly impacted the mental health of today’s teens and tweens.

Haidt presents four key proposals to address this crisis:

1. Delay smartphones until high school (allowing “dumb phones” or smartwatches instead)

2. Implement phone-free school environments

3. Restrict social media use until age 16

4. Encourage more free play to build responsibility and independence

Here are some of my favorite quotes grouped by category

Social Media

While the reward-seeking parts of the brain mature earlier, the frontal cortex-essential for self-control, delay of gratification, and resistance to temptation-is not up to full capacity until the mid 20s, and preteens are at a particularly vulnerable point in development. As they begin puberty, they are often socially insecure, easily swayed by peer pressure, and easily lured by any activity that seems to offer social validation

A fourth trend began just a few years later, and it hit girls much harder than boys: the increased prevalence of posting images of oneself, after smartphones added front-facing cameras (2010) and Facebook acquired Instagram (2012), boosting its popularity. This greatly expanded the number of adolescents posting carefully curated photos and videos of their lives for their peers and strangers, not just to see, but to judge.

the four foundational harms of the new phone-based childhood that damage boys and girls of all ages: social deprivation, sleep deprivation, attention fragmentation, and addiction.

Social media therefore harmed the social lives even of students who stayed away from it. (My added context: students felt left out if they weren’t on a social media app)

Compared with boys, when girls go onto social media, they are subjected to more severe and constant judgments about their looks and their bodies, and they’re confronted with beauty standards that are further out of reach.

Free play

Children can only learn how to not get hurt in situations where it is possible to get hurt, such as wrestling with a friend, having a pretend sword fight, or negotiating with another child to enjoy a seesaw when a failed negotiation can lead to pain in one’s posterior, as well as embarrassment. When parents, teachers, and coaches get involved, it becomes less free, less playful, and less beneficial. Adults usually can’t stop themselves from directing and protecting.

A key feature of free play is that mistakes are generally not very costly. Everyone is clumsy at first, and everyone makes mistakes every day. Gradually, from trial and error, and with direct feedback from playmates elementary school students become ready to take on the greater social complexity of middle school. It’s not homework that gets them ready, nor is it classes on handling their emotions. Such adult-led lessons may provide useful information, but information doesn’t do much to shape a developing brain. Play does.

Experience, not information, is the key to emotional development. It is in unsupervised, child-led play where children best learn to tolerate bruises, handle their emotions, read other children’s emotions, take turns, resolve conflicts, and play fair. Children are intrinsically motivated to acquire these skills because they want to be included in the playgroup and keep the fun going.

The human brain contains two subsystems that put it into two common modes: discover mode (for approaching opportunities) and defend mode (for defending against threats). Young people born after 1995 are more likely to be stuck in defend mode, compared to those born earlier. They are on permanent alert for threats, rather than being hungry for new experiences. They are anxious.

Children are most likely to thrive when they have a play-based childhood in the real world. They are less likely to thrive when fearful parenting and a phone-based childhood deprive them of opportunities for growth

Maturity

If a child goes through puberty doing a lot of archery, or painting, or video games, or social media, the activities will cause lasting structural changes in the brain, especially if they are rewarding.

Natural sleep patterns shift during puberty. Teens start to go to bed later, but because their weekday mornings are dictated by school start times, they can’t sleep later. Rather, most teens just get less sleep than their brains and bodies need. This is a shame because sleep is vital for good performance in school and life, particularly during puberty, when the brain is rewiring itself even faster than it did in the years before puberty.

Friendships

All know that they will be chosen or passed over based in part on their appearance. But for adolescent girls, the stakes are higher because a girl’s social standing is usually more closely tied to her beauty and sex appeal than is the case for boys.

The happiest girls “aren’t the ones who have the most friendships but the ones who have strong, supportive friendships, even if that means having a single terrific friend.”

read more

If you need to create a file in a Lambda you need to write the file to /tmp because it is otherwise a read-only file system. But if you’re emailing a file there’s no need to write the file to the file system, with some minor alterations you can speed up the process and keep the file only in memory.

Current code

csv_file = 'your_file.csv'
with open(csv_file, 'w', newline='') as file:
    writer = csv.DictWriter(file, fieldnames=headers)
    writer.writeheader()
    for item in my_data:
        writer.writerow(item)

    return csv_file

New code

buffer = io.StringIO()
writer = csv.DictWriter(buffer, fieldnames=headers)
writer.writeheader()
for item in my_data:
    writer.writerow(item)

return buffer

and also a small tweak to your email script

Current code

attachment = MIMEBase('application', 'octet-stream')
attachment.set_payload(open(csv_file, 'rb').read())

New code

attachment = MIMEBase('application', 'octet-stream')
attachment.set_payload(csv_file.getvalue())

read more

With the release of LlamaIndex v0.10 imports have changed from top level llama_index package to llama_index.core, llama_index.embeddings, and llama_index.llms

ServiceContext has also been deprecated and replaced with Settings. A concise version of existing code is below

from llama_index import ServiceContext
from llama_index.embeddings import AzureOpenAIEmbedding
from llama_index.evaluation import FaithfulnessEvaluator, RelevancyEvaluator
from llama_index.llms import AzureOpenAI

def evaluate_llama(dataset):
    llm = AzureOpenAI()
    embed_model = AzureOpenAIEmbedding()
    service_context = ServiceContext.from_defaults(llm=llm, embed_model=embed_model)

    faithfulness_gpt4 = FaithfulnessEvaluator(service_context=service_context)
    relevancy_gpt4 = RelevancyEvaluator(service_context=service_context)

    from llama_index.evaluation import BatchEvalRunner

Updated code removes creating and passing ServiceContext around with the new Settings object, which also reduces passing around llmb and embed_model as well. This part is all straightforward, but the migration tool does not take into account needing to add some new packages to requirements.txt

pip install llama_index_core llama-index-embeddings-azure-openai llama-index-llms-azure-openai

Once you’ve installed new packages, you should be able to update your imports. A concise version of the changes is listed below.

from llama_index.core import Settings
from llama_index.core.evaluation import FaithfulnessEvaluator, RelevancyEvaluator
from llama_index.embeddings.azure_openai import AzureOpenAIEmbedding
from llama_index.llms.azure_openai import AzureOpenAI

def evaluate_llama(dataset):
    Settings.llm = AzureOpenAI()
    Settings.embed_model = AzureOpenAIEmbedding()

    faithfulness_gpt4 = FaithfulnessEvaluator()
    relevancy_gpt4 = RelevancyEvaluator()

    from llama_index.core.evaluation import BatchEvalRunner

read more

To squash all git commits on a branch you can run

git reset $(git merge-base master $(git branch --show-current))

There are other required steps, such as ensuring you’re up to date from main, but the gist if what you need is the singular command above

read more

We recently switched out our lambda build image to a debian based image and started receiving errors around glibc.

[ERROR] Runtime.ImportModuleError. Unable to import module 'app':
/lib64/lib.so.6: version 'GLIBC_2.28' not found
(required by /var/task/cryptography/hazmat/bidnings/_rust.abi3.so)

After some googling we realized pip chooses the correct wheel for us and since we were running pip on a different machine than we were running our Python program on, we needed to let pip know about this.

RHEL/CentOS are using manylinux2014 which is what we need to pass to pip

--platform manylinux2014_x86_64

Additionally we do not want to use source packages, so we had to pass

 --only-binary=:all:

Our final command ended up being

python3 -m pip install --platform manylinux2014_x86_64 --only-binary=:all: -r requirements.txt

read more