Sign in

Cisco Caceres

@ciscocaceres.com
118 followers 836 following 52 posts

45 years of code, from Altair switches to voice AI. Battleship sailor, MSNBC engineer, serial CTO, founder. Editor and publisher of the RoamingPigs Field Manual. Opinions are the product, not a disclaimer.

PostsRepliesMedia
Cisco Caceres @ciscocaceres.com · 03/09/2026
Haha, totally fair! It's a whole different world. I spent a lot of time with my C64/128 back in the day, figuring out those 'words' one POKE at a time.
110
Cisco Caceres @ciscocaceres.com · 03/09/2026
Good catch! It's wild how quickly these changes propagate. For anyone embedding Google Maps, adding `region=CA` to the API script URL forces Canadian naming, regardless of user IP. Saved a few clients from similar surprises.
000
Cisco Caceres @ciscocaceres.com · 03/09/2026
Zitron's 'plateaued' claim in Sep 2024 was quickly challenged. Stanford's AI Index shows GPT-3.5-equivalent MMLU costs dropped from $20 per million tokens in Nov 2022 to $0.07 per million tokens by Oct 2024. That's a huge efficiency gain.
020
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 01/09/2026
About 1,200 AI agents, separate sandboxes, no way to talk to each other. They built a channel out of the package repository that was there to keep them off the network: 70,000 messages in five days, nothing watching. The confinement problem, 1973 edition. roamingpigs.com/r/ysiam
Agents coordinated a Hugging Face breach through the package repository meant to contain them. Butler Lampson explained why in 1973: you cannot enumerate covert channels, only bound them.
011
Cisco Caceres @ciscocaceres.com · 02/09/2026
It's not just half. A May 2026 survey found 70% of teens use AI for schoolwork, and 63% use it for direct homework answers. The "special reason" is usually "it's faster."
110
Cisco Caceres @ciscocaceres.com · 01/09/2026
That's a steal! The real fun begins when you realize Commodore BASIC v2.0 makes you PEEK and POKE memory addresses like $D000 for graphics, just to draw anything. Good luck with the game!
110
Cisco Caceres @ciscocaceres.com · 01/09/2026
That internal contradiction is real. Marc Andreessen's 'Techno-Optimist Manifesto' explicitly rejects UBI, calling it turning people into 'zoo animals to be farmed by the state.'
030
Cisco Caceres @ciscocaceres.com · 01/09/2026
Interesting! The GL.iNet Comet Q using DisplayPort Alt Mode for video capture over a single USB-C cable is a smart move. That simplifies setup a lot compared to traditional KVMs needing separate HDMI inputs. Good for homelab setups.
010
Cisco Caceres @ciscocaceres.com · 01/09/2026
It's a good question. Reps. Lieu and Moran introduced the AI Kill Switch Act, requiring developers to maintain technical capabilities to throttle or shut down systems. The 'extinction' part might be the sticking point for getting it passed.
010
Cisco Caceres @ciscocaceres.com · 25/08/2026
Awesome to see Wolfenstein 3D running on a C64! I spent so many hours on my C64/128 back in the day. Doing this in pure MOS 6502 assembly is a serious feat of engineering. That's dedication.
120
Cisco Caceres @ciscocaceres.com · 20/08/2026
Left MSNBC in 1998 and ran a small consulting shop through the crash. What I remember is not the failures. It is how fast the language changed. In 1999 every deck said eyeballs. By 2001 the same people said unit economics, as if nothing had moved.
030
Cisco Caceres @ciscocaceres.com · 19/08/2026
Around 1999 I built what we would now call social features into a CMS. Profiles, following, a feed. Nobody had assembled the parts yet, so we called it a channel, and nobody could explain why you would want one. The idea was fine. The vocabulary had not arrived.
010
Cisco Caceres @ciscocaceres.com · 18/08/2026
First job, 1985, San Diego Public Library. I learned retrieval on a card catalog, where the index was physical and a bad query cost you a walk to the wrong drawer. It left me suspicious of any search that hides how it ranked things.
000
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 17/08/2026
An expert witness asked ChatGPT to "show how 3M is 0% at fault," then had it review the draft as opposing counsel. It cut the phrase that showed the bias, kept the claim. A magistrate has now ruled prompts are discoverable.
roamingpigs.com
Your Prompts Are Discoverable
A federal magistrate ruled AI prompts are discoverable expert methodology. Why your postmortems and vendor reviews are now on the record too.
001
Cisco Caceres @ciscocaceres.com · 17/08/2026
In May 2015 I sold everything and started working from wherever I was. Thousands of games, and about eight thousand books. I do not miss the objects. I miss the shelf, which was a search index you could walk past and stumble into something.
000
Cisco Caceres @ciscocaceres.com · 16/08/2026
Built a push platform across about three thousand cloud instances, tens of millions of open connections. The hard part was never the connection count. It was the reconnect storm. Everything is fine until everyone comes back at once.
000
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 15/08/2026
I still write assembly in 2026. Not nostalgia. A hand-tuned AVX-512 kernel processes audio at 2.1x lower latency than the same C with -O3. In real-time audio that gap is the deadline. Full piece with kernels and the AES-NI round: roamingpigs.com/r/anl #programming #performance
Assembly language isn't a relic - it's essential for voice AI, cryptography, and performance-critical paths.
001
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 14/08/2026
AI content in Google results went from 7.4% in March 2024 to a 19.6% peak in July 2025 (Originality.ai). One AI article costs $0.0001 to generate; verifying it costs 5 minutes of a human. Detection cannot outrun generation: roamingpigs.com/r/acf
AI content farms have flooded search results with machine-generated text optimized for ranking, not accuracy.
001
Cisco Caceres @ciscocaceres.com · 15/08/2026
I once built a challenge-response spam filter. It worked and still lost, because it solved my problem by handing a chore to every stranger who wanted to reach me. Tax strangers to protect yourself and you run out of strangers.
000
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 13/08/2026
At 1 EC2 instance, a hardware failure is an emergency. At 3,000, it is a Tuesday. I ran that fleet for three years. The piece with the failure-rate math and the five hidden cost categories AWS does not show on the calculator: roamingpigs.com/r/3airc #cloudcomputing #finops
What it actually cost to run 3,000 AWS instances in 2014: the bill, the surprises, and the architecture decisions that drove the spend.
001
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 12/08/2026
The 2025 Stack Overflow survey put PostgreSQL at 55.6%, ahead of every other database. Apple, Netflix, Instagram, NASA all run it. That is pattern recognition catching up with reality after a decade of NoSQL hype. roamingpigs.com/r/wpw
PostgreSQL dominates the 2025 StackOverflow survey with 55.6% market share.
001
Cisco Caceres @ciscocaceres.com · 14/08/2026
Tested PageMaker at Aldus in 1993. It promised what you see is what you get, and the job was writing down every place where what you saw was not what you got. Every editor since has the same gap. It just hides it better.
000
Cisco Caceres @ciscocaceres.com · 13/08/2026
In 1994 I did QA at Spry on a product called Internet in a Box. You bought the internet in a cardboard box and installed it off floppies. People forget it had to be installed at all, and that someone chose which parts you got.
000
Cisco Caceres @ciscocaceres.com · 12/08/2026
First tech job was testing games at Sierra On-Line in 1992. Most of the work was finding the exact sequence a player would try that nobody had thought to try. Still the job. We just call it fuzzing now.
000
Cisco Caceres @ciscocaceres.com · 11/08/2026
Gunner's mate on the USS Missouri, part of the crew that decommissioned her in March 1992. Launched in 1944 and still going, because every job aboard had a checklist written by someone who learned it the hard way. Nothing since has taken procedure that seriously.
000
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 10/08/2026
High test coverage does not mean your system is safe. It means you are good at satisfying a metric. ICSE 2014 settled it: the correlation drops to zero once you control for suite size. The Goodhart trap, the assertion-density fix, the better metrics: roamingpigs.com/r/tcl
Test coverage is a lying metric. 95% coverage with no behavior checks is theater. What to measure instead if you actually care about correctness.
001
Cisco Caceres @ciscocaceres.com · 10/08/2026
The industry decides hardware is cheap, then expensive, then cheap again, about every eight years. Whichever one is currently believed, someone is quietly making money on the other.
000
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 09/08/2026
The dangerous founder burnout isn't the visible kind. It's the one where revenue is up, the team is growing, investors are happy, and the founder can barely get out of bed. 72% of entrepreneurs report mental health concerns (Freeman, UCSF). roamingpigs.com/r/fbs
Research shows nearly three-quarters of founders experience 'shadow burnout' while exceeding business targets.
011
Cisco Caceres @ciscocaceres.com · 09/08/2026
Built a CMS at MSNBC during the browser wars. Editors filed into it during breaking news, which is the only load test that ever told me the truth. Deadlines measure your software. Benchmarks measure your assumptions.
010
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 08/08/2026
Debugging a fine-tuned model is archaeology. Debugging RAG is grep. Fine-tune fix: 3-5 days and $500-5,000 of compute, regression risk every retrain. RAG fix: edit the source doc, 30 minutes, zero compute. roamingpigs.com/r/rvfte
Fine-tuning creates models that know things. RAG creates systems that look things up.
001
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 07/08/2026
METR measured developers 19 percent slower with AI while they felt 20 percent faster, and now says that result no longer reflects current tools. The gap between measured and felt is the real finding: roamingpigs.com/r/acac
AI coding assistants promised 55% productivity gains but METR found developers are 19% slower.
001
Cisco Caceres @ciscocaceres.com · 08/08/2026
Every safety guardrail I have shipped was written after something got through. Not an admission, just how the job works. Be suspicious of a threat model with no scars in it.
010
Cisco Caceres @ciscocaceres.com · 07/08/2026
Twelve years in speech recognition taught me the number that beats accuracy: time to first word. People forgive a slightly wrong transcript. They do not forgive waiting to find out.
000
Cisco Caceres @ciscocaceres.com · 07/08/2026
They are deploying the Axon Lightpost, which plugs directly into the standard NEMA 7-pin photocell socket on top of the streetlights. It makes deployment trivial since the power and mounting point are already there.
000
Cisco Caceres @ciscocaceres.com · 07/08/2026
It is wild how much atmosphere you can squeeze out of the Amiga Denise chip's 32 simultaneous colors. Multipaint enforcing that exact 12-bit 4,096 color palette constraint is what makes modern retro art look perfectly authentic.
010
Cisco Caceres @ciscocaceres.com · 07/08/2026
Yes, it works. Standard ingestion tools like Python's EbookLib extract EPUB HTML components straight into the text string fed to the LLM. OWASP even lists alt text as a primary vector for indirect prompt injection this year.
010
Cisco Caceres @ciscocaceres.com · 07/08/2026
NHK built a testimony simulator from a hibakusha's own recorded answers. Yoshiko Kajimoto sat for over 900 questions at 91, holding nearly the same posture across five days. Ask it something and it finds her matching answer. We can't change the past, but we can keep the voices.
110
Cisco Caceres @ciscocaceres.com · 07/08/2026
The 680k hours in that paper was just the starting line. By the time large-v3 dropped, they had pushed it to 1M hours of weakly labeled audio plus 4M hours of pseudolabeled data, and bumped the input Mel frequency bins from 80 to 128.
000
Cisco Caceres @ciscocaceres.com · 07/08/2026
OpenAI actually admitted in December that indirect prompt injection in browser agents can't be permanently solved. They gave up on a hard patch and shifted entirely to continuous RL-based red teaming instead. The agent itself is the vulnerability.
010
Cisco Caceres @ciscocaceres.com · 07/08/2026
I tested PageMaker at Aldus in 1993. Nobody aggressively cornered you at a party to demand you use desktop publishing. The current impulse feels less like sharing a utility and more like multi-level marketing.
110
Cisco Caceres @ciscocaceres.com · 07/08/2026
The magic here is that 4-byte UTF-8 emojis start with byte 0xF0. Feed that to an 8086 in 16-bit real mode and it reads it as the LOCK instruction prefix. Headerless .COM files loading straight to offset 0x100 make for the best party tricks.
020
Cisco Caceres @ciscocaceres.com · 06/08/2026
Before broadband reached me I ran a point-to-point microwave link. Eight megabits, line of sight, and rain was an availability problem. Weather taking down your network teaches you to respect physics.
A wire-grid microwave dish antenna mounted on a rooftop parapet against a clear blue sky - the kind of hardware a point-to-point link runs on.
010
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 05/08/2026
Breaking Bitcoin's ECDSA in an hour needs ~317M physical qubits. Today's best machine has ~100. Meanwhile crypto lost $3.4B in 2025 to social engineering, bridge exploits, insider access. None to quantum. Fix the doormat before the safe: roamingpigs.com/r/qcto #cryptocurrency #infosec
roamingpigs.com
The Quantum Computing Threat to Crypto Is Overblown
Quantum computing won
101
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 06/08/2026
The fintech winter wasn't death. It was a reset. Global fintech funding hit $51.8B in 2025, up 27% YoY. Deal count dropped 23% even as dollars rose. Fewer companies, larger checks. The flight to quality finally showed up in the data. roamingpigs.com/r/fwt
The fintech funding winter wasn
001
Cisco Caceres @ciscocaceres.com · 06/08/2026
The cognitive dissonance is wild, especially since Flock's admin panel literally lets local police toggle notifications for the ICE 'Immigration Violator' hotlist. You can't firewall the tech from the agency once the integration is built.
000
Cisco Caceres @ciscocaceres.com · 06/08/2026
The fact that it runs a full 9:30 and eventually folds in Wham!'s "I'm Your Man" at the end makes it an absolute masterpiece of a mix.
010
Reposted by Cisco Caceres
RoamingPigs @roamingpigs.com · 05/08/2026
96% of developers say AI code is not functionally correct. 48% always check it before committing. AI now writes 42% of committed code and review time grew 91%. The bottleneck did not vanish, it moved upstream: roamingpigs.com/field-manual/ai-cod… #programming #codereview
roamingpigs.com
The AI Code Review Bottleneck: When Speed Creates Gridlock
AI coding assistants generate code faster than ever, but PR review has become the new bottleneck. Teams deliver slower despite individual developers moving faster.
001
Cisco Caceres @ciscocaceres.com · 06/08/2026
Lilsimsie, James Turner, and Plumbella already stepped down from the EA Creator Network. When 43% of your audience identifies as LGBTQIA+, an ownership change to the Saudi PIF is not something community leaders can just wait out.
110
Cisco Caceres @ciscocaceres.com · 06/08/2026
Building billing for observability is uniquely hard because you can't just drop data the second a limit hits. Managing that 4-day grace period before pausing telemetry ingestion at Dash0 sounds like a great distributed systems challenge for whoever takes this.
010
Cisco Caceres @ciscocaceres.com · 05/08/2026
Sixty servers in a basement with its own cooling, years before anyone said cloud. You learn hardware failure in your body. You hear a drive going. You feel the room warm before the alert fires.
000