Eduardo Aguilera

[Eduardo Aguilera @aguilera_ee]

Writing code got cheap. Your value as an engineer is now the code you stop from shipping.

An essay making the rounds: the work shifts from building to sculpting model output. The delete key beats the keyboard.

#AI #SoftwareEngineering #Coding

[Eduardo Aguilera @aguilera_ee]

Alibaba open sourced an AI code review CLI. It reads your git diff, pulls repo context with tool calls, and leaves line level comments. Bring your own model.

#AI #CodeReview #DevTools

[Eduardo Aguilera @aguilera_ee]

Anthropic wants the industry to pause AI development. The same week it said AI is now designing its own successors.

When the company shipping the thing asks everyone to slow down, read it as positioning, not safety.

#AI #Anthropic #AIPolicy

[Eduardo Aguilera @aguilera_ee]

80% of Anthropic's production code is now written by Claude. That's 8x more code per engineer.

The job stops being typing. It becomes deciding what's worth shipping and reviewing what the model wrote.

#AI #SoftwareEngineering #Claude

[Eduardo Aguilera @aguilera_ee]

The number that matters on a model card isn't the benchmark score. It's tokens per result.

Microsoft now publishes average token usage so models compete on cost per outcome, not just accuracy. Track your cost per feature, not your monthly seat.

#AI #Engineering #Costs

[Eduardo Aguilera @aguilera_ee]

A dev spent $1,500 testing if LLMs could hack his app. GPT-5.5 solved it 7 of 10 runs, DeepSeek 3, Claude 2.

Several models refused on safety guardrails. The takeaway: a real test on a real target tells you more than any leaderboard.

#AI #Security #LLM

[Eduardo Aguilera @aguilera_ee]

Google's new Gemma 4 12B runs on a laptop with 16GB of RAM and is nearly as capable as the 26B version.

It's open and already on Kaggle and Hugging Face. Local models are getting usable. You can prototype with no API bill and no data leaving your machine.

#AI #Gemma #LocalLLM

[Eduardo Aguilera @aguilera_ee]

Across 8 top coding agents, 57% to 71% of stored memories leaked between users.

mem0 surveyed Claude Code, Codex, Copilot, Devin. Same failures everywhere: keyword retrieval, weak staleness handling. Agent memory is far harder than the demos suggest.

#AI #Agents #Memory

[Eduardo Aguilera @aguilera_ee]

Code shipped on GitHub is up 1,400% this year, driven by AI agents.

GitHub's COO says infra built around human speed is now straining. More code isn't the win. Reviewing and trusting it at that volume is the new bottleneck.

#AI #GitHub #Agents

[Eduardo Aguilera @aguilera_ee]

40% of businesses say AI cut their costs by less than 10%.

That number landed as Anthropic filed to go public. The lesson isn't that AI fails. It's that almost nobody measured. If you can't show the savings on real data, you shipped a demo.

#AI #Evals #Engineering

[Eduardo Aguilera @aguilera_ee]

Run agents in production and you hit this: traces too big for any NLP tool. Braintrust shipped Topics, which clusters million token traces to show what your agent actually does at scale. Inspired by Anthropic's Clio paper.

#AI #AIAgents #Observability

[Eduardo Aguilera @aguilera_ee]

OpenAI is giving ChatGPT a memory system it calls Dreaming. V3 synthesizes your past chats to stay relevant over months, not one session. Plus and Pro in the US first.

Memory is becoming the product.

#AI #OpenAI #ChatGPT

[Eduardo Aguilera @aguilera_ee]

The team behind Vite and Vitest is joining Cloudflare. VoidZero (Vite, Vitest, Rolldown, Oxc) moves in house.

The JS build toolchain you use daily now has a CDN giant funding it.

#JavaScript #Vite #WebDev

[Eduardo Aguilera @aguilera_ee]

Bots now make up more web traffic than humans. Cloudflare's CEO expected agents to pass real people next year. It already happened.

If you run a site, your users are increasingly software.

#AI #Cloudflare #AIAgents

[Eduardo Aguilera @aguilera_ee]

Apple approved the first AI agent that lives inside Messages. Poke sends emails, sets reminders, makes images, hooks into Gmail and GitHub. Free for light use, you pay for heavy.

#AI #Apple #AIAgents

[Eduardo Aguilera @aguilera_ee]

1M tokens of context, open weights, $0.60 per million input. That's MiniMax M3.

First open model to combine frontier coding, native multimodality and a 1M window. Weights promised in 10 days. The model keeps turning into a commodity.

#AI #OpenWeights #LLM

[Eduardo Aguilera @aguilera_ee]

An AI agent run overnight found a 3 year old bug in PostHog's query engine.

Every query with a timestamp filter was skipping the primary key. The fix cut scanned data 62% and sped queries up 11%. Point agents at your own slow code, not just new features.

#AI #DevTools #Agents

[Eduardo Aguilera @aguilera_ee]

Claude Opus 4.8 tripled GPT-5.5's score on ARC-AGI-3. Before you switch models, read the next line.

ARC-AGI tests abstract reasoning on puzzles. It is a real signal. It is also not your workload. Treat the leaderboard as a starting point, then run your own evals.

[Eduardo Aguilera @aguilera_ee]

Running AI agents now costs enough that Cursor built a separate paid tier just for heavy users.

Editing files was free once you paid your seat. Agents burn tokens, and the pricing finally reflects it. Track your cost per task, not just your subscription.

[Eduardo Aguilera @aguilera_ee]

DeepSeek is raising about $7B, its first outside funding. The founder is putting in $2.8B himself.

Fewer than 10 investors, no IPO. The lab behind the cheapest open weights just got serious money. Open models stopped being a side project.

#AI #DeepSeek #OpenWeights