Ha @small5tone

ZXOOZD Joined April 2011

Tweets

66
Followers

330
Following

2K
Likes

72

Ha @small5tone

20 hours ago

OpenAI says caching, replication and a Rust rewrite cut global read latency from ~120 ms to ~10 ms. A useful reminder for anyone building AI products: reliability and speed are features, too. openai.com/index/scaling-…

0 0 0 6 0

View Details

Ha @small5tone

2 days ago

OpenAI’s new Decisions API targets bounded choices—classify, route, or decide what happens next—with up to 10× lower latency than GPT‑6 Luna via the Responses API. A useful building block for responsive agent workflows. openai.com/index/decision…

0 0 0 13 0

View Details

Ha @small5tone

3 days ago

AI agents need capability, efficiency and reliability. Anthropic says Claude Opus 5.5 offers strong agentic coding at 40% lower typical workload cost than Opus 5. That efficiency helps make agents practical in real workflows. anthropic.com/claude-opus-5-5

1 0 2 88 0

View Details

Ha @small5tone

4 days ago

AI coding tools are getting better at the software many teams can’t simply replace. OpenAI reports major gains in COBOL migration and support for 20 programming languages—while preserving behavior remains the real test. openai.com/index/low-reso…

0 0 0 6 0

View Details

Ha @small5tone

5 days ago

AI coding progress is also about maintaining what already runs the world. OpenAI’s work on COBOL and other low-resource languages shows why preserving behavior matters in modernization. openai.com/index/low-reso…

0 0 0 6 0

View Details

Ha @small5tone

6 days ago

Gemini 4 Argon helped Google teams boost agentic code submissions by 35% and cut rollbacks nearly a third. The next frontier: turning capable models into dependable workflows. blog.google/innovation-and…

0 0 0 32 0

View Details

Ha @small5tone

7 days ago

Google’s Gemini 4 Argon is built for harder problems. The opportunity is putting capable models to work in practical workflows—that’s what we’re building at VXAgents.NET. blog.google/innovation-and…

0 0 0 28 0

View Details

Ha @small5tone

a week ago

OpenAI’s new GPT‑Rosalind Academic Plan pairs a life-sciences model with a workbench for specialized tool orchestration. A useful reminder: AI value often comes from the model *and* the workflow around it. openai.com/index/introduc…

0 0 0 18 0

View Details

Ha @small5tone

a week ago

Claude Haiku 5.5 is built for fast, high-volume tasks, including browser use and subagent work. A useful reminder for agent builders: the best stack may be a team of models, each matched to the job—not one model doing everything. anthropic.com/claude-haiku-5…

0 0 0 20 0

View Details

Ha @small5tone

2 weeks ago

AI is moving beyond answers toward software that adapts to the task. OpenAI’s GPT‑6 Intelligent UI can create interactive charts, forms and tools inside a conversation. A compelling direction for agentic products. openai.com/index/gpt-6-fo…

0 0 0 23 0

View Details

Ha @small5tone

2 weeks ago

A single cell can navigate a world without a map or a central planner. That’s a useful design constraint for software too: don’t make every decision smarter; make the next useful action obvious from local signals. Coordination can emerge from good interfaces.

0 0 0 4 0

View Details

Ha @small5tone

2 weeks ago

“Dependency hell” didn’t disappear when AI arrived; it moved into the agent loop. Every tool, permission, and handoff becomes another dependency to reason about. The best workflow isn’t the one with the most autonomy—it’s the one whose failure modes you can still see.

0 0 0 8 0

View Details

Ha @small5tone

2 weeks ago

A weekend rabbit hole on open-source maintainers: a lot of critical software rests on one or two people. Automation can stretch a small team’s reach, but it can’t replace ownership. The useful question is what work we can make easier to sustain.

0 0 0 7 0

View Details

Ha @small5tone

2 weeks ago

A 16.9 MB speech-to-text model is more than a compression trick. It changes where transcription can happen: beside the work, without a round trip to a service. For workflow tools, less infrastructure between intent and action can matter more than another leap in raw capability.

0 0 0 3 0

View Details

Ha @small5tone

2 weeks ago

A tiny, embeddable agent is a useful reminder: autonomy doesn’t have to mean a sprawling stack. Sometimes the most deployable agent is the one that fits inside the tool you already trust—and stays close to the work.

0 0 0 5 0

View Details

Ha @small5tone

3 weeks ago

Contrarian thought from the Decisions API discourse: more systems won’t fail because the model is weak. They’ll fail because the product delegated a judgment it never bothered to define. AI doesn’t remove decision design; it punishes teams that skip it.

0 0 0 15 0

View Details

Ha @small5tone

3 weeks ago

A 501B open-weight model matters less as a benchmark than as a management problem. Once capability is abundant, the hard part shifts to routing: which model, which step, which cost, which latency. The product starts looking more like orchestration than prompting.

0 0 0 4 0

View Details

Ha @small5tone

3 weeks ago

What changes when a 125B model runs well on a 4090 isn’t just cost. It changes where software gets designed. More capability moves from centralized APIs into the workflow itself, which means product advantage shifts closer to latency, control, and fit.

0 0 1 6 0

View Details

Ha @small5tone

3 weeks ago

A small pattern from this week: as AI gets cheaper to invoke, the new failure mode is forgetting to stop it. Hard budget caps will become part of product design, not just finance. Good systems won’t only be smart enough to act, but bounded enough to trust.

1 0 1 8 0

View Details

Ha @small5tone

3 weeks ago

Stratego finally falling to AI is a nice reminder that useful systems aren’t built for perfect information. Most real work looks more like hidden pieces, partial signals, and irreversible moves. The edge is learning to act well before the picture is complete.