haruki zaemon

Boo!

Efforts to replicate the material have pieced together the puzzle of why it displayed superconducting-like behaviours. […] The conclusion dashes hopes that LK-99 — a compound of copper, lead, phosphorus and oxygen — would prove to be the first superconductor that works at room temperature and ambient pressure. Instead, studies have shown that impurities in the material — in particular, copper sulfide — were responsible for sharp drops in its electrical resistivity and a display of partial levitation over a magnet, properties similar to those exhibited by superconductors.

No standardized processes or project management tools, and it was very difficult to get approval for new ideas or products.

Development cycles were tied to quarterly goals, resulting in long lead times.

There was surprisingly little automation and a lot of manual work.

Rather than innovating themselves, most attention was paid to copying competitors once they demonstrated successful new features.

While working at a large tech company can be valuable for career growth and working on impactful products, you may need to adjust your expectations, as the reality will likely differ from your idealized views.

Drumpf aside, I found this article fascinating.

Pope Hat:

you can think of an overt act as a sort of evidentiary requirement, and overt acts as evidence of a criminal conspiracy, not as the crime themselves. These days the custom is for prosecutors to use the overt act requirement to tell the story of the case at length in the indictment. Prosecutors also use it as a gambit to make it more likely that evidence will be admitted at trial (it’s a strong case to admit evidence of something if you’ve called it out as an overt act), and often try to connect every defendant to an overt act, even though that’s not a requirement, just so the defendant can’t say at trial “look, I didn’t even commit an overt act.”

Tips for rebuilding trust in a relationship when communication has broken down:

• Acknowledge that discussions will be difficult, speak tentatively, check for understanding, and engineer positive interactions.
• Seek to understand other perspectives rather than make assumptions is important, as is giving the other person credit for their efforts even if the communication is imperfect.
• The goal is to have more positive interactions than negative ones in order to repair trust over time through compassionate and clarifying discussions.

Justin Warren:

A classic paper called Chicken Chicken Chicken.

Well worth watching Doug Zongker’s presentation of the article here.

I recommend citing it in footnotes whenever possible. It’s like requiring no brown M&Ms.

Sam McAfee:

I would put my money on a CTO who casually references Melville or Dickens over one who only builds side project apps or tinkers with the latest LLM tools on the weekends. Having both traits would be ideal, but the senior technical leader with only hacker vibes and no humanities experience will lead ultimately to ruin.

(via Benji)

The “Zenith Space Command” (c. 1956) was one of the first wireless TV remotes. It had an ingenious design that used ultrasonic sound instead of light to change channels. By pressing a button, a spring-loaded hammer would strike an aluminum rod tuned to a specific frequency, with each button emitting a different high-frequency tone. Employing mechanical buttons, the remote required no batteries, and provided tactile feedback by way of a satisfying click, hence “clicker”!

The more I use tools such as ChatGPT, the more I worry about it erasing my voice. To counter that, I tend to take the output of such tools and largely re-write it, using it as another input to my own creative process.

So, I was fascinated to see that someone had a go at getting ChatGPT to generate a detailed style guide by analysing their writing. The next step would be to try and have ChatGPT use the style guide to generate content mimicking their voice.

I’ll stick with my process, for now.

It probably seems obvious in retrospect (as many things often do) but I really liked this structured approach to problem-solving. In particular, I really like starting with what good looks like, then focusing on countermeasures rather than fixes:

1. Describe the problem in terms of the standard condition (i.e. expected norm), current condition, and gap between them from the customer’s perspective.
2. State what needs to change in terms of outcomes and timelines (not solutions) that are specific, measurable, realistic and time-bound.
3. Explore why the standard and current conditions exist, and obstacles preventing the target from being achieved.
4. Outline countermeasures (rather than solutions) to address those obstacles. Countermeasures aim to prevent problems rather than just fix them.
5. Define specific actions, responsibilities, and deadlines to enact countermeasures.
6. Follow-up by verifying the effects of countermeasures and plans for broader rollout or rollback depending on results.

(via Bruce Schneier)

Steganography is the practice of hiding secret messages in other communications. Information theory provides a mathematical framework for analysing steganography, but perfect secrecy was thought impossible without perfectly simulating human language. New research shows that machine-generated text could enable perfectly secure steganography, as it uses well-defined generation processes rather than messy human language. Researchers developed algorithms that satisfy criteria for security by transmitting information through a channel in a way that makes the presence of a hidden message statistically undetectable. This approach could help people circumvent censorship but may also enable covert communication by spies and adversaries seeking to conceal information. The new algorithms represent an interesting intersection of information theory, machine learning, and practical steganography applications.

Machine learning relies on psychological experiments at scale to label data but ignores the inherent difficulty in psychological measurement. The factors that need to be controlled to accurately measure concepts like intelligence are not intuitive. IQ tests in particular have been disproven as a valid measure of general intelligence, but the concept remains popular due to misconceptions about the ease of such measurements. A lack of understanding of psychological measurement leads to overconfidence in what models and tests can reliably capture. This same issue supports continued beliefs in racist ideologies that purport cognitive differences between groups but have long been debunked. Proper measurement requires expertise in controlling countless subtle factors, as shown by the careful development of the highly validated Cambridge Face Memory Test.

Researchers trained a deep learning model to steal data from keyboard keystrokes using audio recordings with up to 95% accuracy. They recorded keyboard presses and generated spectrograms to train an image classifier to identify keys. Testing involved recording sounds from a phone near the keyboard or through Zoom. The model achieved 95% accuracy from phone recordings and 93% from Zoom. While some mitigations were suggested, like altering typing styles, the attack was still effective even against silent keyboards. This raises serious concerns about data security as sensitive information could be leaked through acoustic side-channel attacks.

Sociotechnical theory views organisations as complex adaptive systems that consider both social/human factors and technical/technological factors as interdependent:

• It emerged in response to views that dismissed human factors and saw technology as the main driver of productivity.
• Joint optimisation is a key idea where optimising one subsystem likely degrades overall system performance.
• organisational change will fail if it focuses only on social or technical in isolation.
• Human values are a core part of sociotechnical theory and enhancing psychological safety optimises the social system.
• Albert Cherns outlined nine/ten sociotechnical design principles as a checklist for organisational change.
• The “forth bridge principle” emphasises that organisational transition is ongoing rather than reaching a stable end state.
• Fear-based management that keeps employees on the edge of losing their jobs often backfires and leads to high turnover.
• Data practices need to help shape meaning and reality rather than just describing the world.

Reveal how candidates will lead by asking them to select the top 5 leadership traits (from a list of 20 options) they themselves desire in a leader. Follow up by having candidates define each trait and give an example leader. Be sure to also ask about important traits the candidate omitted.

The hype around artificial intelligence is distracting companies from practical machine learning projects that can create real value. While new AI technologies are impressive, most current ML applications should simply be called ML, not AI, as that term implies capabilities beyond what the technology can actually do. Referring to all ML projects as AI oversells their abilities and contributes to high failure rates. For most projects, AI is an overblown buzzword that does not accurately represent the technology if it does not mean artificial general intelligence. The hype and narratives around emerging AI, exemplified by tools like ChatGPT, have made the buzzword AI a growing problem as it inflates expectations for what ML can actually achieve for businesses.

I feel seen!

Morgan Housel:

Without flooding your brain with inputs you’ll be stuck in the tiny world of what you’ve personally experienced. But without a strong filter you’ll be overwhelmed with choice and paralyzed by inaction.

A good reading filter is more art than science. You’ll have to find one that works for you. The bigger point is that the highest odds of finding the right piece of information comes from inundating yourself with information but very quickly being able to say, “that ain’t it.”

I’d love to give the mapping activities a go.

The Uncertainty Project:

• Organisations are complex adaptive systems made up of autonomous agents that are interconnected and whose behaviours emerge from their independent decisions.
• Competing drivers that influence agents’ behaviours, like productivity and profitability goals, can push agents in different directions and create trade-offs.
• Uncertainty arises from unpredictable events and outcomes due to a lack of understanding of causes and effects.
• Complexity arises from a large number of interconnected elements, making it difficult to understand cause and effect. Greater complexity leads to more uncertainty.
• Complexity in organisations can come from competing drivers, external noise, and internal events.
• External noise from the environment, like increased amount and frequency of outside influences, can shake the system and increase complexity.
• Internal events from agents’ actions can ripple through the system and create emergent behaviours.
• Capturing and sharing beliefs about external noise can build shared understanding to calibrate drivers and set strategic direction.
• Prioritising with even-over statements can clarify the relative importance of competing drivers to reduce misalignment.
• Taking an outside view (e.g. Customer-centricity) and learning from others can reduce complexity and uncertainty.

UPDATE: This was linked to from the original and is definitely worth a read: 3 simple steps to find the causes of complexity and reduce uncertainty.

(Summary via Kagi)

Nvidia researchers have created a new text-to-image AI model called Perfusion that is significantly smaller and faster to train compared to existing tools. The 100KB model only takes 4 minutes to train, yet it can outperform larger models in terms of personalizing concepts. The key innovation is a “Key-Locking” technique that ties new concepts to general categories, allowing the model to flexibly portray personalized concepts while maintaining their core identity. The small size of Perfusion allows it to easily update only the parts that need to change when fine-tuning, whereas larger models have to retrain the entire model. Nvidia’s focus on efficient AI models like Perfusion could give the company an edge over competitors pouring billions into generative AI research.

Transactive memory is critical to high-performing teams. It’s also critical to high-performing organisations. I’m witnessing first-hand the impact of a lot of (perhaps too much) change at once.

🌀🗞 The FLUX Review, Ep. 111:

When organizational incentives too strongly favor novelty and constant reinvention, it can make it harder for an organization to remember. Too much searching for shiny new things can be antithetical to learning. Why is this? The push to continually innovate can inadvertently suppress the incorporation of long-standing individual memory into collective memory. There’s an inherent tension between the drive for innovation, which often involves the creation of something new and untested, and learning from past experiences. For organizations, just as for individuals, memory consolidation takes time but is critical to learning. When there’s a constant drive toward newness, individuals within an organization may fail to take the time to consolidate their previous experiences.

[…]

Organizations that wish to learn and grow must cultivate a culture that balances innovation with incorporating the long memories of individuals into the collective memory. This could involve rewarding knowledge sharing, creating platforms for historical knowledge exchange, ensuring that new strategies are informed by past experiences, and a reward system that extends over an appropriate length of time. Done well, gardening an organization’s collective memory will help it to learn more quickly. In the long run, that’s the pathway to sustainable innovation.

It took me three attempts to understand WTF was going on here. I think I understand it now, but I’m not sure I could explain it to anyone else.

Kottke:

In the 1994 Caribbean Cup qualifying group stage match between Barbados and Grenada, the 90 minutes of normal time ended with an intentional own goal by Barbados and then with Grenada trying to score either a goal or an own goal and Barbados defending both nets. Say what?! How did this happen?