Samuel Fajreldines

🌐

[Samuel Fajreldines @devindolar]

@LLMJunky They won’t do that. If they do, gpt will bring a competitor and win market share

[Samuel Fajreldines @devindolar]

@bridgemindai I doubt they will remove fable from subscription for one simple reason. If they do, OpenAI will lunch a competitor in their subs and win the market.

[Samuel Fajreldines @devindolar]

30min on Mythos

[Peter Steinberger 🦞 @steipete]

Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore.

You should be designing loops that prompt your agents.

[Samuel Fajreldines @devindolar]

@DavidOndrej1 I’d rather to have security to use my bank apps. iOS still #1 for it

[Samuel Fajreldines @devindolar]

@cjzafir Apple spent trillions of dollars and couldn’t do that. Awesome work, congratulations!! 👏🏼👏🏼

[CJ Zafir @cjzafir]

Here's a teaser of our Mac-1 model.

> 6.6B model
> runs locally (on any Mac)
> requires 7GB RAM (12GB ideal)
> can use 487 MacOS native tools
> perform multi-tool chained tasks
> reasoning: ON
> output: ~65 tok/s

We built a robust application layer around the model to make UI/UX MacOS native. The "model-focused" SaaS era is here.

Stay tuned for more.

[diva @divaagurlxw]

As an AI Engineer. Please learn

>Harness engineering, not just prompt engineering

>Context engineering, not just long prompts

>Prompt caching vs. semantic caching tradeoffs

>KV cache management, eviction, reuse, and memory pressure at scale

>Prefill vs. decode latency and why they optimize differently

>Continuous batching, paged attention, and throughput optimization

>Speculative decoding vs. quantization vs. distillation tradeoffs

>INT8, INT4, FP8, AWQ, GPTQ, and when quantization hurts quality

>Structured output failures, schema validation, repair loops, and fallback chains

>Function calling reliability, tool contracts, argument validation, and idempotency

>Agent guardrails, loop budgets, tool budgets, and termination conditions

>Model routing, graceful fallback logic, and degraded-mode UX

>RAG architecture: chunking, embeddings, hybrid search, reranking, and freshness

>Retrieval evals: recall, precision, grounding, attribution, and citation quality

>Evals: golden sets, regression tests, adversarial tests, LLM-as-judge, and human evals

>LLM observability as a first-class discipline: traces, spans, tokens, latency, errors, and drift

>Cost attribution per feature, workflow, tenant, and user journey not just per model

>Safety engineering: prompt injection defense, data leakage prevention, and permission boundaries

>Multi-tenant isolation, cache safety, and cross-user context contamination prevention

>Fine-tuning vs. in-context learning vs. RAG vs. distillation and when each is the wrong tool

>Latency, quality, cost, and reliability tradeoffs across the full inference stack

>Production failure modes: hallucinated tool calls, malformed JSON, stale retrieval, runaway agents, and silent eval regressions

[Samuel Fajreldines @devindolar]

@TeksEdge OMG. Testing!

[Samuel Fajreldines @devindolar]

@leaks_legit And m7 will outperform m6. This is life.

[Samuel Fajreldines @devindolar]

@Its_Nova1012 The same that makes a CEO unreplaceable: To know what to ask.

[Samuel Fajreldines @devindolar]

@Im_IrushiK Marketing

[Samuel Fajreldines @devindolar]

@Xi_fak3 Because windows is just horrible for software development.

[Daniel Lougen @DJLougen]

A 2.1GB model on my gaming PC CPU just beat a $10M AI model on HumanEval.

Here's exactly how:

The model: Qwen2.5-Coder-3B-Instruct — 3.1B params from Alibaba, quantized to 4-bit.
Downloaded in 30 seconds.

The hardware: Intel i9-12900K. No GPU. A $350 consumer CPU.

The score: 89.0% (146/164 problems passed)

Cohere Command A+: 218B parameters, $10M+ training cost, requires 2x H100 GPUs. Scored 75%.

We're +14 points. On a cpu.

I resurrected busyBeaver:
→ Prompt engineering (expert coder framing)
→ pass@3 retry at 3 temperatures (pushes 80% → 89%)
→ Code extraction from markdown output
→ Sandboxed test execution (15s timeout, crash recovery)
→ Checkpointing (resume from any crash point)

The model writes the code. The harness measures it fairly.

Eval protocol: Textbook standard. Feed signature + docstring → generate code → run tests → count passes. No tricks. No benchmark training. No contamination.

Honest scorecard:
- HumanEval: 89% vs 75% ✅
- MBPP: 70% vs 72%
- MMLU-Pro: 27% vs 68% ❌ Expected (code model vs knowledge model)

You don't need $10M to beat a $10M benchmark. You need a 2GB model + a clean eval harness + a gaming PC.

Code: https://t.co/48mhZP7reY

[Samuel Fajreldines @devindolar]

@cjzafir I'm doing exactly the same! Let's connect

[Ramin Nasibov @RaminNasibov]

Training a new AI model

[Samuel Fajreldines @devindolar]

@cjzafir Im using deepseek to generate datasets

[Samuel Fajreldines @devindolar]

@DJLougen Poor little Beaver :(

[Samuel Fajreldines @devindolar]

@cjzafir How did it make your internet faster? 🤔 Did you mean your network or your internet?

[Samuel Fajreldines @devindolar]

@thsottiaux @kr0der Something that let me use codex 24/7 without limits and in spark speed