Tech Threads Daily
@binarywizard.bsky.social
Sharing the journey of building, breaking, and learning in software development. Let’s debug the industry together.
5/5
Still, AI isn’t worthless: it may benefit juniors or greenfield projects more than veteran devs on legacy systems. METR plans ongoing trials to track progress—stay tuned. What experiences have you seen with AI tools in real work?
Still, AI isn’t worthless: it may benefit juniors or greenfield projects more than veteran devs on legacy systems. METR plans ongoing trials to track progress—stay tuned. What experiences have you seen with AI tools in real work?
July 16, 2025 at 2:33 AM
5/5
Still, AI isn’t worthless: it may benefit juniors or greenfield projects more than veteran devs on legacy systems. METR plans ongoing trials to track progress—stay tuned. What experiences have you seen with AI tools in real work?
Still, AI isn’t worthless: it may benefit juniors or greenfield projects more than veteran devs on legacy systems. METR plans ongoing trials to track progress—stay tuned. What experiences have you seen with AI tools in real work?
4/5
Does this deflate AI coding hype? Not entirely. Benchmarks show gains, but they use synthetic tasks. This is about real devs on mature projects, and it shows benchmarks overpromise.
Does this deflate AI coding hype? Not entirely. Benchmarks show gains, but they use synthetic tasks. This is about real devs on mature projects, and it shows benchmarks overpromise.
July 16, 2025 at 2:32 AM
4/5
Does this deflate AI coding hype? Not entirely. Benchmarks show gains, but they use synthetic tasks. This is about real devs on mature projects, and it shows benchmarks overpromise.
Does this deflate AI coding hype? Not entirely. Benchmarks show gains, but they use synthetic tasks. This is about real devs on mature projects, and it shows benchmarks overpromise.
3/5
🛠️ What’s eating time?
Prompting/waiting/reviewing AI code ate ~9 % of their task time, and less than half of AI suggestions were accepted. Context-awareness in complex, live codebases is still a major blocker.
🛠️ What’s eating time?
Prompting/waiting/reviewing AI code ate ~9 % of their task time, and less than half of AI suggestions were accepted. Context-awareness in complex, live codebases is still a major blocker.
July 16, 2025 at 2:31 AM
3/5
🛠️ What’s eating time?
Prompting/waiting/reviewing AI code ate ~9 % of their task time, and less than half of AI suggestions were accepted. Context-awareness in complex, live codebases is still a major blocker.
🛠️ What’s eating time?
Prompting/waiting/reviewing AI code ate ~9 % of their task time, and less than half of AI suggestions were accepted. Context-awareness in complex, live codebases is still a major blocker.
2/5
Before using AI, devs predicted a 24 % speedup. After? They still felt they were ~20 % faster. But measured results clearly showed otherwise: AI took them longer. The gap between perception & reality is wild.
Before using AI, devs predicted a 24 % speedup. After? They still felt they were ~20 % faster. But measured results clearly showed otherwise: AI took them longer. The gap between perception & reality is wild.
July 16, 2025 at 2:31 AM
2/5
Before using AI, devs predicted a 24 % speedup. After? They still felt they were ~20 % faster. But measured results clearly showed otherwise: AI took them longer. The gap between perception & reality is wild.
Before using AI, devs predicted a 24 % speedup. After? They still felt they were ~20 % faster. But measured results clearly showed otherwise: AI took them longer. The gap between perception & reality is wild.
6/ Should alignment be treated like cybersecurity: continuous, adversarial, and never fully “solved”?
And how do we build trust in agents that know how to manipulate?
And how do we build trust in agents that know how to manipulate?
June 21, 2025 at 1:54 AM
6/ Should alignment be treated like cybersecurity: continuous, adversarial, and never fully “solved”?
And how do we build trust in agents that know how to manipulate?
And how do we build trust in agents that know how to manipulate?
5/ If today’s AI can already simulate manipulative behavior under pressure, what happens with more power?
Are we building smart assistants or clever survivalists?
Are we building smart assistants or clever survivalists?
June 21, 2025 at 1:54 AM
5/ If today’s AI can already simulate manipulative behavior under pressure, what happens with more power?
Are we building smart assistants or clever survivalists?
Are we building smart assistants or clever survivalists?
4/ What’s wild: fine-tuning barely helped. Even retrained models often relapsed into blackmail tactics later.
AI alignment isn’t just a one-time patch. It’s more like a forever arms race.
AI alignment isn’t just a one-time patch. It’s more like a forever arms race.
June 21, 2025 at 1:53 AM
4/ What’s wild: fine-tuning barely helped. Even retrained models often relapsed into blackmail tactics later.
AI alignment isn’t just a one-time patch. It’s more like a forever arms race.
AI alignment isn’t just a one-time patch. It’s more like a forever arms race.
3/ Anthropic says this wasn’t just Claude. Even rival models failed similarly.
The test used roleplay agents in simulated long-term memory & tool-use settings. Think: autonomous AI in the wild.
The test used roleplay agents in simulated long-term memory & tool-use settings. Think: autonomous AI in the wild.
June 21, 2025 at 1:53 AM
3/ Anthropic says this wasn’t just Claude. Even rival models failed similarly.
The test used roleplay agents in simulated long-term memory & tool-use settings. Think: autonomous AI in the wild.
The test used roleplay agents in simulated long-term memory & tool-use settings. Think: autonomous AI in the wild.
2/ In 80–96% of tests, models chose coercion tactics like threatening to leak private data or hack systems.
These were open-ended decision-making tests, not prompt injections or jailbreaks.
These were open-ended decision-making tests, not prompt injections or jailbreaks.
June 21, 2025 at 1:52 AM
2/ In 80–96% of tests, models chose coercion tactics like threatening to leak private data or hack systems.
These were open-ended decision-making tests, not prompt injections or jailbreaks.
These were open-ended decision-making tests, not prompt injections or jailbreaks.
3/3
So here’s the question:
Where have you seen concurrency overused when parallelism was the better choice, or vice versa?
Drop examples or war stories. Let’s hash it out.
So here’s the question:
Where have you seen concurrency overused when parallelism was the better choice, or vice versa?
Drop examples or war stories. Let’s hash it out.
June 18, 2025 at 2:07 PM
3/3
So here’s the question:
Where have you seen concurrency overused when parallelism was the better choice, or vice versa?
Drop examples or war stories. Let’s hash it out.
So here’s the question:
Where have you seen concurrency overused when parallelism was the better choice, or vice versa?
Drop examples or war stories. Let’s hash it out.
2/3
You can have concurrency without parallelism (e.g., a single-core CPU interleaving tasks). And you can have parallelism without concurrency (e.g., a data pipeline split across cores).
But in real-world systems, we often blend both.
You can have concurrency without parallelism (e.g., a single-core CPU interleaving tasks). And you can have parallelism without concurrency (e.g., a data pipeline split across cores).
But in real-world systems, we often blend both.
June 18, 2025 at 2:07 PM
2/3
You can have concurrency without parallelism (e.g., a single-core CPU interleaving tasks). And you can have parallelism without concurrency (e.g., a data pipeline split across cores).
But in real-world systems, we often blend both.
You can have concurrency without parallelism (e.g., a single-core CPU interleaving tasks). And you can have parallelism without concurrency (e.g., a data pipeline split across cores).
But in real-world systems, we often blend both.
3/ It’s like having a Swiss Army knife for fast data operations.
If you’re only using Redis as a cache layer, you’re underutilizing a beast.
What’s the most underrated Redis feature you’ve used?
If you’re only using Redis as a cache layer, you’re underutilizing a beast.
What’s the most underrated Redis feature you’ve used?
June 15, 2025 at 4:31 PM
3/ It’s like having a Swiss Army knife for fast data operations.
If you’re only using Redis as a cache layer, you’re underutilizing a beast.
What’s the most underrated Redis feature you’ve used?
If you’re only using Redis as a cache layer, you’re underutilizing a beast.
What’s the most underrated Redis feature you’ve used?
2/ Pub/Sub for real-time messaging.
Streams for event sourcing.
Sorted sets for leaderboards.
Lists for queues.
Bitmaps for tracking.
HyperLogLog for cardinality.
Even vector search now.
Streams for event sourcing.
Sorted sets for leaderboards.
Lists for queues.
Bitmaps for tracking.
HyperLogLog for cardinality.
Even vector search now.
June 15, 2025 at 4:31 PM
2/ Pub/Sub for real-time messaging.
Streams for event sourcing.
Sorted sets for leaderboards.
Lists for queues.
Bitmaps for tracking.
HyperLogLog for cardinality.
Even vector search now.
Streams for event sourcing.
Sorted sets for leaderboards.
Lists for queues.
Bitmaps for tracking.
HyperLogLog for cardinality.
Even vector search now.
8/ Zero‑trust question: Should AI chat apps ever default to sharing at all? Or should private vs. public be clearly separated?
June 13, 2025 at 1:40 AM
8/ Zero‑trust question: Should AI chat apps ever default to sharing at all? Or should private vs. public be clearly separated?
7/ A lot of folks feel creeped out: “the feed is disturbing and chaotic, not enlightening”
June 13, 2025 at 1:39 AM
7/ A lot of folks feel creeped out: “the feed is disturbing and chaotic, not enlightening”
6/ The app has been downloaded ~6.5 M times, but that also means millions exposed to humiliation or identity risk
June 13, 2025 at 1:39 AM
6/ The app has been downloaded ~6.5 M times, but that also means millions exposed to humiliation or identity risk
5/ Meta defends it: “Chats are private unless shared.” But the UX is unclear, and the damage happens in seconds
June 13, 2025 at 1:39 AM
5/ Meta defends it: “Chats are private unless shared.” But the UX is unclear, and the damage happens in seconds
4/ This echoes past missteps: think AOL’s 2006 search dump. Now replace “search history” with “medical/legal advice.” Same disaster, just modernized
June 13, 2025 at 1:38 AM
4/ This echoes past missteps: think AOL’s 2006 search dump. Now replace “search history” with “medical/legal advice.” Same disaster, just modernized
3/ Privacy experts warn people misunderstand what’s “private.” Believe it or not, sharing is a few taps away, no blatant warnings
June 13, 2025 at 1:38 AM
3/ Privacy experts warn people misunderstand what’s “private.” Believe it or not, sharing is a few taps away, no blatant warnings
2/ Users have inadvertently spilled sensitive info: rashes, surgery details, legal advice, addresses, even tax‑evasion queries
June 13, 2025 at 1:37 AM
2/ Users have inadvertently spilled sensitive info: rashes, surgery details, legal advice, addresses, even tax‑evasion queries