Johnny Silicio

AI/ML engineer | Built AI for 35M+ users @ Replika | €1.1M raised as Co-Founder

[antirez @antirez]

Calling classifiers "decision models" is a crime against machine learning. Those models decide *less* than LLMs, as you can think the CoT as a form of serial processing of information that such classifiers lack. Those models can be either *small* LLMs that are not let think, but instead the logits at the last position of the prefill are used to categorize in classes or, when they are BERT-alike, they compress the input meaning and project a class: in both cases they decide a lot less than an LLM.

[Jacob Coxon @hilbertspaess]

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

[Johnny Silicio @johnny_silicio]

@levelsio Having more human data is about increasing the knowledge of LLMs, but LLMs already knows everything. Now it's a matter of making the LLM useful on always more complex tasks, and for that they don't more human data, there's RLVR.

And yes the dead internet theory is true

[antirez @antirez]

You will discover more and more that the anti-AI part of the society will be the most aggressive and fascist, especially in the US I believe the extreme polarization of the society could create big issues.

[antirez @antirez]

I worked at Redis till 2020, so the irrelevancy phase was just from 2020 to 2024 (when I rejoined). Also I'm older now. Please could you fix, without the help of AI, your graph to reflect the current time line? Also you may want enjoy my Vector Sets implementation. It is hand-written, a few months ago, and is one of the best HNSW implementations out there.

[Johnny Silicio @johnny_silicio]

@markproduct https://t.co/KFalfh37fp

A mobile app to remotely control agents in your laptop

[Johnny Silicio @johnny_silicio]

@JakeFleshner Mobile Agents

[Johnny Silicio @johnny_silicio]

@dom_gag_96 🍕

[Johnny Silicio @johnny_silicio]

Show HN this week: Hopper, agentic interface for mainframes and COBOL. AI-on-legacy-code is the quietest big market in software. Insurance, banks, government still run on it. https://t.co/0BbbYT1LP5

[Johnny Silicio @johnny_silicio]

Instructure paid ShinyHunters' ransom Monday for Canvas LMS data. 275M user records, 8,800+ institutions. CEO Daly: 'we got the balance wrong'. Education is now an organized-crime line item.

[Johnny Silicio @johnny_silicio]

Show HN this week: Statewright. Visual state machines for AI agents. Pattern is overdue. Most production agents are spaghetti prompt+if/else under the hood. https://t.co/avMPIgIr9t

[Johnny Silicio @johnny_silicio]

Fix: Semantic Kernel Python ≥1.39.4, .NET ≥1.71.0. Real fix: treat every tool registration like a privilege grant. The frontier model isn't your attacker; the next prompt your agent reads is. 4/4

[Johnny Silicio @johnny_silicio]

Microsoft disclosed two RCEs in Semantic Kernel last week. CVE-2026-26030 (Python SDK, CVSS 9.8): one prompt → arbitrary code via the InMemoryVectorStore filter. Prompt injection is now RCE. 1/4

[Johnny Silicio @johnny_silicio]

Pattern: neither CVE needed an exotic jailbreak. One was a filter expression left unsanitized. The other was a method auto-registered as a tool by reflection. The model just had to ask. 3/4

[Johnny Silicio @johnny_silicio]

Kunvar Thaman's reward hacking benchmark on 13 frontier agents. 0% (Sonnet 4.5) to 13.9% (R1-Zero). RL post-training is the dial: V3 0.6%, R1-Zero 13.9%. 72% include explicit CoT. arxiv 2605.02964

[Johnny Silicio @johnny_silicio]

What to watch: whether discovery surfaces other lab decisions where licensing was rejected for cost. Anthropic, OpenAI, Google have similar lineage. The phrase 'standard practice' gets named. 4/4

[Johnny Silicio @johnny_silicio]

Filed May 5: five major publishers and Scott Turow allege Zuckerberg 'personally authorized and actively encouraged' Meta's torrent of 267 TB from LibGen for AI training. Dataset scale matters. 1/4

[Johnny Silicio @johnny_silicio]

The new wedge isn't 'AI trained on pirated data', that's the standard claim. It's 'the CEO personally directed the procurement decision'. Discovery liability moves up the org chart. 3/4